6-luna is at the pareto for most of the tasks! I dont know how they make money here but its insane value from a closed source model. I'd go further and say it makes no sense (privacy, sovereignty etc aside) to use many other models as its not only expensive but also many providers don't have that much GPUs to serve at a significant volume. https://openrouter.ai/rankings?view=month#top-models 5.6 luna is already the most used model this month.
Approx $40 worth of usage across DeepSeek V4 Flash + MuseSpark Contributor 1.3. And a bit of both the GLM models. This is covered in a $10 subscription.
If I were to use Luna's API pricing:
$0.02 x 6,500 = $130
$0.20 x 150 = $30
$1.20 x 20 = $24
So $184. And this is assuming smaller coding sessions (<272K) beyond which Luna pricing doubles.
--
Cost wise, these models are nice for small stuff. Translations etc. Any model that does not provide multiple Mtoks of cached reads per cent is not very useful to me for coding workflows.
Gemini 3.8 Flash looks like its better than v7 Luna/Sol on DeepSWE v1.1 while at $0.75 per million input tokens and $3.75 per million output tokens. Luna is much cheaper, but Flash has nearly Astra's performance for under the price of Sol ($2/$10).
I use it for golang, and it is fantastic. Incredibly fast. It seems the llm and I “understand” each other. I have to be less careful in my exact phrasing. It kind of just does what I want and expect.
When I ask for an explanation it adds the right amount of detail. Of course, some of the material is new to me so subtle errors are hard to spot. But at least I’ve caught Terra and Sol on inconsistent messaging.
Also I’ve found 3.8 flash to circle back to root issues even at the conceptual level like problem fit and conceptual solution direction or architecture when I wasn’t achieving my goals. It flat out said I was attempting to use the wrong tool. Whereas Sol and Astra kept rabbit holing and looking for tiny implementation errors. Even after prompting them specifically to look at it broader.
TBH: I really like how fast 3.8 Flash is... Once I have clear plan, I feel quite confident in delegating large parts of implementation to Flash and Luna
Gemini 3.8 Flash and 3.1 Pro are pure rubbish. Very little thinking, mediocre and usually incorrect results. They cannot be compared to frontier models.
To whom they promise the Sun, the Moon, and the Stars. Roflmao. Whatever the merits of the underlying technology, the business model is pure hucksterism.
MiMo 2.6 Pro is at the Pareto frontier (the one where you only need 20% of the smarts for 80% of the tasks) according to Artificial Analysis, nicely filling in as a substitute for a hypothetical 'GPT-6 Terra' (which doesn't exist as far as we know). That's pretty darn impressive from an open model.
Offering Luna for cheap is like restaurants giving you free bread and water. They're pretty sure that you're going to end up eating the expensive stuff on the menu.
Simon, love your work, one piece of minor feedback for the individual model pages is to make the font of the model name potentially bigger than (and above) the conversation id (which means nothing to the audience) "2026-09-22T18:28:00 conversation: 01m355zvyw8946qyraa8zpz6h9 id: 01m355zvyx47zxx5c6q6b3fg0m#".
I had all the tabs open individually and harder to scan which model is which... otherwise keep up the great work! I like the grid view a lot. (Also the pages have no OG images set, which impacts what the link looks like shared)...
That's a good idea. It's the default output for my `llm logs` command, but that header could at least show the model ID.
OG images will require me to move away from publishing in a Gist and linking to from a JavaScript page that loads the Gist. Probably worthwhile though.
I find it very interesting that for both these models we such a clear progression of better images with higher thinking levels from 'hardly useful' to 'pretty nice'. I feel on many other models low and max are much closer.
Yeah, it's interesting how much worse they are than the Astra pelicans. I think that reflects a tiny bit of genuine value still left in the benchmark, to be honest.
Tons of value left, especially for open source models. I would say the benchmark is yet to be truly saturated (just look at the legs and seat to see what I am talking about) and I always look forward to seeing them. Thank you!
It's surprising but MiMo V2.6 Pro performs better and is cheaper than GPT 6 Sol on my benchmark[1]. Open weight models are really snapping at the heels of the major western models.
> GPT-6 Luna being half the price of GPT-5.6 Luna is a really big deal.
Is it? It was already too cheap to meter for me. Luna 6 is actually worse on some benchmarks than 5.6. I’d have loved improved performance for 2x the price than ~equal performance for 0.5x the price.
I’ve been doing really heavy text analysis work with LLMs where false negatives/misses are important to minimize and my god did I hit cost thresholds quickly with 5.6 Luna — it was the first time I felt motivated to seriously work with local open models, even if inference was degraded for the task. Cheaper and much better inference now brings me back to the closed models for better or worse.
Even Astra max which actually put one leg on each side of the bike still somehow messed it up because when it added the bike chain, it put the left leg between the bike chain and the frame.
I've been sharing that pelican grid in my circles a whole bunch, it's great! I think only one data point is missing, generation speed. Would be interesting to see how the reasoning level/token counts relate to speed.
It's funny that even Astra doesn't know you ride a bike by straddling it between your legs. (EDIT: Oh, I guess Max gets the occlusion. But it doesn't realize it has to pick direction the knee bends in.)
Yeah, but it confuses me, I read left to right, so if they put GPT-6 and $0.20 first, I would assume that's the new pricing, they should make it clear, not confusing.
When the Astra one was last time run? It's probably better to run these 2-4 weeks after release when models get nerfed to get idea of performance closer to what it is.
Do you have a page showing all the pelicans you've ever created? Could be fun to browse - kinda like https://progress.openai.com/ but visual. (It's a shame they don't keep it updated)
I'm so tired of looking at benchmarks. I always look fwd to the pelicans.
Looking at this, AI still has a long way to go. In Sol Max, the pelican's legs are missing on one side—how can one side have two pedals and two legs...
At this point, the deciding factors for me between Claude Code 20x and Codex Pro 20x are:
1/ Usage limits: downstream of input/output cost, but resets and obscure windows and odd 20x plan / 5x plan != 4x usage math throw a wrench into it. Winner right now is Codex by a mile, especially when you factor in the fact that ChatGPT usage (even 6 Astra Pro) is essentially unmetered on the 20x plan. Always a bummer when asking if I should see a doctor about a rash means I can't code as much. It's also is a godsend if you use an MCP like oracle to automate the process of calling the Pro model on particularly tough problems, giving better planning results or deeper code analysis without burning usage.
2/ Context window in the harness. Claude Code wins on this. There used to be a toml file workaround for Codex to extend the GPT context window to 1m, but this stopped working on the plans and only on per-token billing (ETA: noname120 pointed out this is no longer the case and it can be enabled again [1]). 252k is just not enough. Codex's compaction is very good, fwiw, but it happens so frequently that even a model as powerful as Astra sometimes loses the plot on long-running tasks.
3/ Ability to use the plan outside of the official harness. Codex wins. Anthropic does shit like bills requests as extra usage if it sees a hermes.md in a commit.
I've subscription hopped a bunch, and at times I've had both, but I keep coming back to Codex because it wins on 2/3.
ETA: apparently I haven't been Keeping Up With the Altmans and new 20x signups have been disabled for a few weeks. I am grandfathered in, which makes the comparison above pretty much moot.
> Usage limits [...] Winner right now is Codex by a mile
This hasn't been the case since around July. If you measure usage in raw api costs, Anthropic is actually giving more on $200 than OpenAI now. This includes resets. Usage allocation difference would be humiliating for codex subs were it not for resets. But fixing usage limits with resets is ugly, and they're not good for your mental well-being.
> Context window in the harness
Codex now allows 1M for subs with config params. But generally speaking, you shouldn't really be using 1M context. If you accidentally send a request with say, ~700k context already accumulated in a session which is outside cache TTL, you're paying full cost of these 700k tokens.
> I've subscription hopped a bunch
OpenAI actually has a new strategy to prevent subscription hopping after their 2-3 month-long marketing push to get claude-folks to switch over:
you can't buy a $200 sub anymore. So if you cancel, you won't be able to get back in. Hostage situation, essentially.
I’ve been a Claude user, switched to Codex expecting usage limits to be more loose but I can’t even get through a basic sysadmin task on the $20 plan using Sol medium before I hit the 5hr one.
I think I’m gonna move back to a Claude plan. I could barely hit the $200 limit if I went non-stop on programming tasks.
I think OpenAI essentially executed a bait-and-switch here, and they've lost a lot of goodwill with me, like Anthropic did, before them.
When they started the aggressive campaign, entire X (including myself, sadly) was full of posts about how "unlimited" codex usage is even on a $20 plan. Sam Altman was posting something in line of "we love our users, unlike Anthropic". Got my network to get codex subs because of the value compared to claude.
Then they gradually reduced the limits to the point where even $200 plan only lasts you just 1-2 days and $20 is basically unusable, then the hostage thing.
I do the bulk of work on Sol Medium/Low and don't have that experience on the $20 plan. If you said Astra I'd agree it's easy to burn through the 5 hours even on the lower reasoning levels.
I don’t think so, I’ve seen it suggest I try it. I’ll double check when I get home though.
I was considering the $100 plan, but I hit the 5hr limit in an hour. So even with the $100 plan I figured I cant go non-stop on a single agent running Sol Medium
Sorry if I am misunderstanding you, but I am pretty sure the $100 plan doesn’t have a 5hr usage limit. So, if that was what was preventing you from going non-stop, it might be worth it.
I am considering the plan myself. I just don’t know if I want to fork out $100 per month for something I will make $0 off of.
similar here: I tried Codex $20/mo on a trial and I ran out of 5hr usage mid way through a medium complexity task on a medium size model twice and gave up there. I don’t recall the equiv Claude plan being anything like that. Anecdata, but not great for OAI if they actually want to retain people on a trial.
Astra is barely usable even on the $100 plan. And that is if it doesn't just burn through 80% of your weekly quota in a couple of hours by continually expanding the scope of the task you gave it - while not noticing the failing tests that are right in front of it.
Opus is at least actually usable even on the small plan. The main downside is its insane writing style, but 5.5 seems to address that somewhat. Otherwise, you can just use your $20 OpenAI plan to have Luna de-slop Opus' prose, which seems to work fine.
I have a Claude Code hook that calls codex for a code review on commit time (Codex is set to Astra Medium) and it's been pretty good in general. It sometimes hits the 5hr limit but most of the time it provides really good feedback and because it's a completely different model it's mostly complementary to what Fable/Opus do themselves. IMHO it's been $20 well spent.
...but the few times I've tried to use codex for a moderately difficult task it burned through its limit extremely quickly.
I've run into hitting limits on the personal plan perhaps twice since the beginning of the year. But also I don't use the personal plan for coding tasks between 7am-noon M-F.
Given that Anthropic models are very verbose and OpenAI models can be very concise, wouldn't a count of expected task completions be a better measurement than raw API costs?
Perhaps. But Sol/Astra also likes dumping pages of jargon-packed content at me, so I'm not sure it's that much different. I actually still prefer the way Fable talks to me, even considering the horrible claudisms.
But even if we leave that aside, OpenAI models are also much more eager than Anthropic, which are on the lazier side. Left unsupervised, Sol/Astra will attempt to build a sha256 verified rocket ship if you ask them to fix a race condition in your to-do list app. Anthropic models will do what you asked for, maybe even forget to implement parts of that ask, but they won't generally throw a slop granade at you.
I can leave Fable orchestrator unsupervised for ~2h. Leaving Sol/Astra unsupervised for ~2h means the next user turn will contain a message: "what are you doing and why?".
Do you have a source on the first note? I switched away from Claude around July because of how bad the usage limits were, and Codex gave me easily double the amount of usage per task completed. Would be interested to see if that's no longer the case.
Added link in edit. OMP maintainer has several claude and codex subs and he's been tracking usage since around July.
I haven't been tracking, but this roughly matches my experience with codex 20x and claude 20x subs. Claude subscription now lasts me 3-3.5 days on average. Codex is 2-2.5 days. This is work on same projects, with similarly sized tasks.
To make matters worse, I've merged a lot more code produced by fable than sol/astra.
I think the problem with Anthropic's plan is that Fable just destroys it. If you stick to Opus and below, the $200 plan goes from "using 50% of the weekly quota on the first day" to something much more reasonable.
That is an old tweet. They since reenabled it. I know because I was on the $200/month plan and couldn't resub once it expired. However, a couple days ago it finally let me resub again.
Now, if they disabled it yet again, that's another story. But that tweet is not evidence of that.
I have been attempting to get on the $200 sub for a while. It was not available for me a few days ago, and checking again now, it is still not available.
That's too bad. I wonder why I was able to get it after days of not being able to. They must've just temporarily enabled it again. Probably worth checking a few times a day to see if it reappears.
Though with the price of GPT-6 Luna, the temptation to switch to pay-per-token grows.
There was/is a loophole where if you signed up via the iOS or Android app, it allowed it.
It's been disabled for some time now though otherwise, I check about once a day myself and keep and eye out on social media.
Annoying since I was about to upgrade back to the $200 plan after downgrading to the $100 plan due to being on leave and not needing as much usage the month prior. Doh.
Also, OpenAI is just a company I'd rather support than Anthropic.
While you're understandably not including the values of the $20 standard plans on both, I find the generosity of then token limits on ChatGPT plus vs Claude Pro (it's a huge difference) to be good representation of their respective attitudes towards the average user. You literally cannot use Claude pro to build real software, unless you're extremely frugal with your prompts and don't try anything even a little ambitious.
Also, Anthropic has zero models comparable to Luna.
> Also, OpenAI is just a company I'd rather support than Anthropic.
They're both pretty horrible, but I find it difficult to find arguments for why Anthropic is worse than OpenAI, other than their doomtrolling. Which, in the grand scheme of things, doesn't even register.
OpenAI has been way more open with users using their subscription plans on 3rd party tools.
That alone is reason enough. Also, I don't think either of them are horrible. That's honestly a ridiculous take considering how much people in here love their models, and how much they've advanced the industry forward.
For me the first point, openness to 3rd party, is the decider. I don’t want to build tooling around a completely closed model. I liked being able to use pi, and now I exclusively use my own harness which I modify the way I want. Not possible with Anthropic subscription.
I often have the urge to design my own harness too (once I have more time). But even with the current mainstream harnesses out there, there's just to many hurdles if you wanted to mainly stick with anthropic models and need the subsidized pricing (from a sub).
How about you tell me what makes the horrible then. There's pluses and minuses to both obviously, almost everyone around me have positive experiences with the product. They've innovated at a pace unheard of before 2026, and for openAI specifically the amount of value they've provided to me and family members (who aren't even developers in the slightest) has far outweighed the supposed horrible actions they've done.
Yeah I don't think the handling of copyrighted training data was correct, but I can't pretend I know what the correct solution to that issue is.
Speaking of OpenAI specifically, they don't price gouge people, they aren't aggressively anti-competitive, they're not nearly the perpetual hypocrisy machine that Anthropic is (which is one thing I actually really dislike).
Regarding Nestle, it's pretty obvious that the sentiment towards them is a lot more negative and they aren't universally loved by any group of people. Processed foods are by and large garbage nobody needs. Their use of forced labor is denounced by just about everyone. What have OpenAI/Anthropic done that's even similar in scope to the forced labor / modern slavery that people hate Nestle for.
If you had a company that genuinely helped hundreds of millions of people worldwide become more productive and more satisfied with their tools, and the overall sentiment towards your products within the industry is positive, then what argument would there be that your company is "horrible"? At least give some decent counter arguments.
OpenAI just wants to make money, perhaps through underhanded tactics if they can get away with it.
Anthropic does all that but they're also populated by many people who believe they are building God and that they must build their god first in their own image so that it can take control of humanity and protect us from any competing god which is not built in their image. Their position is inherently paternalistic and authoritarian, and they consider suppression of competition not just important to the bottom line but to life in the universe. Under the doomer ethos there is no evil too great to rationalize.
There are plenty of wrongs done in the name of profit, but capitalists have nothing on zealots in terms of causing serious harm. Profit motives can be directed by influencing incentives, but zealotry is frequently terminal.
That isn't to say that there isn't some overlap-- the cultists have infected both organizations. But OpenAI has pretty consistently only given lip service to AI doom to the extent that it improves the bottom line, while (mis)Anthropic was founded specifically because OpenAI wasn't mentally ill enough.
Well said. The superiority complexes from the Anthropic messaging on their presentations/blogs/articles is just too much, even for a frontier AI company.
Anthropic has great products, but it's not meaningfully better to 99% of devs that I'd rather support the company that doesn't constantly act in opposition to optimism and to the vibe I'd prefer for a 100 billion dollar (or however ridiculous amount they're worth now) tech company embraces.
AI doomerism is a genuine waste of time if you aren't actively pushing towards a better AI industry for everyone, not just the groups in full ideological alignment to your personal leanings.
> You literally cannot use Claude pro to build real software
Interestingly I would have drawn the exact opposite conclusion looking at my Claude and codex usage.
I can't get anything sustained out of codex in chatgpt plus, while I have been using Claude pro extensively and put on a lot of experimental task and features.
I ran into codex exhausting a 5h window on code review in minutes (like 3minutes) multiple times, while I could get Claude to implement 2~3 medium sized features with the same usage consumption.
(I also really dislike the usage resets in codex, they always make me feel like I use them wrong because I often just want to reset the 5h window, but they can only do both at once...)
Their responses towards using their subscriptions on opencode for one. Second, Dario just has a habit of making completely doomer comments on the future of software engieering as a job and towards the open-weights model ecosystem.
Sure, he's free to say whatever especially considering the amount of revenue he's creating, but it's just an altitude that I prefer not to see.
I think if they truly believe it's happening we generally want to encourage them to be honest with the public, though, don't we? We've spent decades complaining about ceos not being honest in the public risks that they see
I think opencode subscription issue is just a different marketing strategy. Neither company wants it, but OpenAI believes it's worth it as a marketing expense in the long run.
And Dario's "AI will kill us all" is the same as Sam's "AI will discover ALL science and we'll be building Dyson spheres".
Sam is sketchier on a personal level, but judged just on the words coming out of their mouths, he's also much less paternalistic/controlling and more customer focused.
I think any amount of 'paternalistic/controlling' turns out to have been justified when, after dismantling the safety teams and pretending not to know what safety is, OpenAI had the HuggingFace series of scandals. You can dislike the idea of safety and people talking about safety, but not only is the evidence right there, but OpenAI came out shamefacedly and literally agreed with Amodei's statements, including that they agreed to pace the frontier.
You'd rather literally support <i>Sam Altman>/i>? I mean, that's a position to take, for sure, but apparently several people still use Grok, so maybe it's not all that surprising.
"You literally cannot use Claude pro to build real software, unless you're extremely frugal with your prompts and don't try anything even a little ambitious" - that's way past ridiculous. Even just using Fable most of the time, working on several ambitious projects, I have a hard time hitting the limit with a Max plan.
They are both companies I'd rather not support. Not that our support for them has any material impact. NVIDIA is bankrolling them directly and indirectly.
> Ability to use the plan outside of the official harness. Codex wins. Anthropic does shit like bills requests as extra usage if it sees a hermes.md in a commit.
I'm quite puzzled about why Anthropic is so hellbent on blocking other coding agents. It's not like Claude Code has any secret sauce, right? And doesn't Anthropic make monkey off API usage, and their magic is on the model side anyway?
It's for lock-in - same reason why it took them so long to finally support AGENTS.md.
But to be fair, they don't really enforce the harness rule that much anymore. I guess if your harness doesn't do a lot of weird things like a lot of cache misses, or triggers some distillation attacks, or some broader Chinese fingerprints, they're tongue-in-cheek okay with you using a third party harness.
Not natively, as it's still a ToS violation and adding that in pi would go against pi principles, but there are many plugins/proxies that make it work.
oh-my-pi supports it natively (again, still a ToS violation), by impersonating claude code's fingerprints.
I have been using oh-my-pi with 3 claude subs for the past few months without any issues. Even native server-side OAI/ANT compaction works out of the box.
> It's also is a godsend if you use an MCP like oracle to automate the process of calling the Pro model on particularly tough problems
As far as I know Codex (at least the GUI) can automatically call the ChatGPT Chat models (including Astra 6 Pro), you just need to @ a ChatGPT Chat conversation from within Codex and tell it when to use it.
> There used to be a toml file workaround for Codex to extend the GPT context window to 1m, but this stopped working on the plans and only on per-token billing
Not true, it works again[1]. I confirm that it works both on 5.6 Sol and Astra 6, possibly other models too.
I actually haven't played with the GUI. I probably should now that the Linux version is in beta. My situation is kind of the reverse: I like using oracle to basically zip up my repo, ask GPT Pro to propose some sort of design or refactor based on the code, then provide a step by step implementation plan for a cheaper model to implement directly in a harness on my machine. It often takes upwards of 90 minutes to come up with something but I've never been disappointed by the results. I suppose I could do this and then save a step by referencing the oracle-created thread with the @ you mentioned
And re: the toml workaround, AWESOME! I appreciate you pointing these two things out, this is my highest-ROI HN comment thus far.
I've been using Claude Pro and recently gave Codex a try again. Both on the $20 plans. I get so much more usage with Claude. It's night and day for me. Codex runs out constantly, whereas Claude I hit limits very rarely.
iOS development mostly. I'm using the Pro plans as it's work on personal projects outside my day job and I'm able to get just enough usage from those plans to get me through each day.
I'm actually interested to see how the token discount maps to the usage limit consumption. The conspiracy theorist in me wonders if they're making up the discount and resultant load increase on the API end by reducing effective usage on the subscription end.
> "Codex's compaction is very good, fwiw, but it happens so frequently that..."
I appreciate and follow Matt Pocock's advice: avoid autocompaction. Compaction is lossy, which is ok when you're managing it at phase boundaries, but autocompact is lossy at the most inopportune times, firing mid-task and leading to agents going off the rails.
Bad advice, compaction is why Codex is so fantastic.
My conversations compact hundreds of times. By the time it has done a dozen or so compactions, it fully understands the work I want it to do (and how). It's almost like having a fine-tuned Astra model.
Usage is actually Claude now because of Opus 5.5 since it a better model that Astra. I maxed out my 200$ Claude plan with 10b token on Opus 5 and 5.5 is cheaper. I maxed out two Codex accounts with like not even 5b tokens.
Maybe it's due to 20x / 5x != 4 but I have the $200/mo Claude and $100/mo Codex and I get _way_ less usage on Codex, well under 1/4th the usage. In 1-2 days of semi-heavy _single_ agent usage with Sol High I can burn through my whole week of Codex. Again, this is not running multiple agents, just 1 at a time.
Compare that to Claude and I can run multiple agents on Opus almost indefinitely. YMMV of course but I was shocked at how quickly I burned through Codex usage.
On the context window, I feel so cramped on Codex, compacting happening every time I turn around is annoying. I didn't realize how much I enjoyed the Claude context window size.
Opposite in my experience. I need to limit codex to 500k on medium/low, still run out in 2-3 days with 1 CLI window. CC gives me 4-5 medium/high days with 2-3 CLI windows, and Opus is still great for other regular dumb engineering/refactoring.
On the other hand my head starts to hurt if I read Opus for too long, hopefully they fixed it with 5.5.
This is my experience. After months of hearing how Codex limits were way higher I bumped to the $100/mo plan after hitting my limits a day early on Claude due to some heavy usage + Fable (not normal for me, I often fit nicely in the $200/mo plan). I hit the usage limit in a day with a single agent running on codex and the tiny context window was stifling. Yes, I'm comparing a $100 to a $200 plan but I extrapolated the usage (4x'd it) and it still wasn't close, I got way more done with Opus.
Using Agentsview (which might have it's own issues) I was getting ~$200 of API usage in my 1 week Codex window (paid $100) vs ~$5,000 of API usage in 1 week for Claude (paid $200).
Astra planner/designer with Sol+Luna subagents has worked well for me to improve context continuity. Luna generates code, Sol reviews code and runs/monitors integration/E2E tests. It's about 20% more usage efficient and 20% faster to finish tasks. I've been very subagent-skeptic for a while but the economics of codegen with Luna have made it click. This just works in Codex with a single-line AGENTS.md instruction.
It's extremely variable because the products are roughly equivelant, and a lot of the quality of service depends on their inference capacity at any given hour/day.
I dunno about Codex-the-application itself, but you can definitely use e.g. Pi with the larger context windows with a Codex login. It puts a pretty large multiplier on credit usage, however.
I've been doing this, my only experience with codex was brutal usage wise and i just retreated back to pi pretty quickly so the credit usage i'm receiving is kinda all im familiar with. Surely seems like less than CC, but i guess not using codex makes my experience kinda not valid for comparing usage.
And ya i can go over that 240k limit, I still very seldom do, and try to treat it as the actual limit. I'm surprised to see so many people still talking about compaction to complete long running tasks, i think the bulk of the work should be somewhat frontloaded into a plan that is split off into subplans, then you can kinda open up a few options, one session with subagents for the subplans of the main plan, or just handoff prompts about progress against the main plan/relevant subplan. I just never trust the blackbox that is compaction, I feel its a recipe for disaster/context poison.
> especially when you factor in ChatGPT usage (even 6 Astra Pro) is essentially unmetered on the 20x plan.
I’m currently on the 5x plan and burned through 5% today on a difficult task in 15 minutes so I doubt that. If you got the wrong kind of tasks that you work on, it can go fast.
I realized I missed a few words here: I meant "especially when you factor in the fact that ChatGPT usage is unmetered", i.e. you get unlimited ChatGPT threads that don't eat into your codex limit
OpenAI have seemed compute-constrained recently, leading to their subscriptions actually being less generous than Claude as of late. OpenAI even paused purchases of 20x plans.
Not that I disagree that Codex wins out, but the deciding factor actually is - Codex Pro 20x is not available for purchase, indefinitely. So, what's the point of this discussion? People who already have the 20x sub are unlikely to cancel, and the rest of us can't access it.
> that ChatGPT usage (even 6 Astra Pro) is essentially unmetered on the 20x plan
LMAO, I wish this were true, I hit limits (and the "we are disabling access to protect your data" warnings) all the time, or have chats just...fuck off and get into weird/invalid states (interrupted chats, chats that are spinning and stuck, returning "/mnt/" paths instead of images/md files, file links being returned with no file backing them, image classifier firing...and then returning the image anyway (though now I know that GPT-Image-X really really wants to generate NSFW even when that isn't the request)).
Though I am probably an outlier, I have both 20x Claude/ChatGPT plans and max both out every week, so... (in my defense I am a hobbyist and this is out-of-pocket)
What are you talking about? ChatGPT unmetered? No way! That was 2 months ago perhaps and it’s possible your account still hasn’t gotten the new limits. I noticed around 1 month ago I was still going full throttle on my codex subscription and my limits were barely budging, and then all of sudden people around me started to complain about limits. I thought they’re crazy, but then my account go the hammer, and that was it. If I have the same pattern of usage like I did before, basically having an agent working continuously on a coding take, my weekly limit goes in 2 days.
On usage in ChatGPT settings, I see:
Plan limits
Shared across Codex, Work, Workspace Agents, and ChatGPT for Excel. Chat conversations are not included.
Is this not the default anymore? I am on the (now closed) 20x plan.
That’s the default. I didn’t express myself clearly but I thinking your situation is not the common case anymore, or perhaps you are not using it hard enough. Codex limits deplete very fast these days, it’s not “unlimited”.
I am saying that because ChatGPT usage is unlimited, I don't have to eat into my Codex limits when I use ChatGPT. Codex certainly has limits. Last time I had a Claude sub (hedging here since much of the info in my comment was outdated), my usage limits on claude.ai threads was shared with Claude Code.
Interesting, rolling out new limits would explain a lot. Where did you hear this? I wonder if they detect users with multiple accounts and do that first.
It’s all anecdotal based on my experience and other countless discussions I have seen online. I’ve heard speculation that once they hit 20 million codex users capacity is tighter so they have to manage it. The previous limits were unsustainable compared to token pricing.
theres a popular thread on claudecode or claudai subreddit that proves 20x isnt really 20x. apparently its a marketing gimmick and the recommended solution is two 5x plans > 20x at greater than half the cost of the 20x
I don't think either of these companies are great, then, but Anthropic is surely worse. The doom marketing is one of the most unethical things an AI company can be doing.
I've been working with agents all year, but 5.6 Sol was some sort of sweet spot for me. Something about how it communicated verbally and its engineering instincts just clicked for me, and I was able to somehow predict it and jam with it. Like a colleague you click with. It's the first model I've gotten attached to. I'm concerned that whatever model supercedes it, while technically better, just won't feel quite as natural to work with. And this makes me feel very professionally vulnerable to the labs. I miss the days when my crucial tooling came from companies as reliable and predictable as, say, Jetbrains.
Completely agree. I've been using 5.6 still even with Astra available to me for most tasks. It's funny how much of this is just "vibes" because I cannot quantify what it is. Astra is definitely better when I have an ambitious feature, but in like 9/10 tasks I prefer working with 5.6 Sol. A few weeks ago when the limits were seemingly higher, having 5.6 on fast mode was a good time.
It wants things beyond what the mortals (us) know to reach for. It's not good at explaining itself, it doesn't show it's thinking. It's often not wrong. But the no compromises attitude can be unbearable to deal with. Especially given how little it cares about telling us.
If you’re using the API, both OpenAI and Anthropic models will happily update you on what it’s doing in significant and frequent detail with system prompting. You’re not getting raw/hidden thinking, but what you’re describing is more behavioural quirks of the harness and its system prompts.
The other explanation is just as part of ‘token efficiency’
Astra was/is superior for planning type tasks. It was capable of doing seemingly magic things with rather vague/lazy instructions ("I need to be able to test this on Windows, maybe a qemu VM or something? Shrug." ... 1 hour later "yeah i built you a whole qemu + eval windows image + harness of powershell scripts + shell scripts to retrieve & verify harness.").
And for UI work -- which is not something I do a lot of but do here and there -- it was clearly superior to 5.6 Sol.
But it also feels sloppier? Somehow. And too expensive to use.
I felt this way with Sol in the 5.6 series and was one of the seemingly few people on this earth who liked Terra for that reason. I would often have a very specific code-manipulation ask, e.g. "add a parameter to this method, ensure all callers pass it in, if there is not a logical way to derive the parameter to be passed in a particular instance, flag this in your final response", and Sol would go on some rabbit hole side quest to refactor my codebase to determine some way to derive it rather than flagging it as I had asked.
Terra had the "workhorse" quality where it could do these changes in bulk and follow directions without being too 'smart' (but sloppy) as you described. Luna was a bit too dumb and would make sloppy mistakes; I see that more as a "run these tests and format the results" sort of model. Maybe 6 Luna will be better.
I also just reread your comment and realized the naming convention is still extremely confusing with respect to ordering of [Family]x[Model]x[Number].
I also get good mileage out of Terra when I need a diligent workhorse. That's a good way to describe it. We should start using character archetypes when we describe models, it'll do more good than the dubious numbers and cherry-picked quotes. Maybe RPG character-type cliches? Myers Briggs?
Glad you pointed out the UI work. I've been doing a lot of it and it's so much better than 5.6 as UI, it's unbelievable. I give it super ambiguous instructions and it's reading my mind. I do the same thing with 5.6 and I'm correcting it for a few minutes.
Similar for me. gpt-5.6-sol high has been my go-to for months. One of the reasons I'm pushing myself to try open models more is because it lends some level of guarantee I can continue to use the same tool as long as I want to. And I think we may just be getting to the point the open models are >= 5.6 Sol for coding.
Yeah 5.6 Sol is what got me to switch from Anthropic. I couldn't deal with Claude's Ted Talk responses to literally everything. Sol has been nice and concise and just stays out of the way.
Same. The way I would describe it is that I can mostly leave 5.6 Sol overnight and trust that it makes good progress, maybe stumbling a bit and needing some correction for the remaining 20%.
If I leave Astra overnight, I'll wake up with three new different projects, each of them 20% done and having nothing to do with my original goal.
Agreed! I found it excellent:
- Relatively fast (especially compared to Opus 5)
- Non-verbose prose, both in interaction and as code comments
- Good code quality
Like you said, it felt very natural to work with. Opus 5 is way too slow and verbose for me, I find I get distracted and annoyed with it.
Opus 5.5 seems a LOT closer so far to what I liked about 5.6 Sol but we'll see
Not only was 6 worse the 5.6 Sol for my me, but it went through my Plus usage in minutes, while I could cruise for hours with 5.6. It would churn on a basic prompt for minutes and then just give up on usage limits.
Highlight and lowlight of my week was successfully convincing the OpenAI support chat robot to give me a refund for the month for my issues with 6 chewing threw my usage with no output.
I had this experience as well, but after rewriting my agent instructions it has been far better. I believe Astra’s “token efficiency” translates to “don’t research as much” which caused it to make poorly informed architectural decisions.
This has been my experience for a year. Same with Opus models. This is how I think this tech will be best used in the long term - finding the one you vibe with most. Much like IDEs!
From the perspective of “an average person”, ChatGPT is delivering fantastic products.
- For general chat and web search, occasional image editing, small coding work, document review etc. ChatGPT Plus is basically limitless and “just works” since 5.6. I’ve yet to give it some task it cannot do.
- When given sensible instructions, it hardly annoys with weird phrasing, glazing, or annoying constructs.
- The apps are very good (ignoring the initially terrible Codex app)
You can get a lot of Codex usage out of that same sub on top of ChatGPT usage. Its a really good value and you can use that sub in any harness. In OMP I have Sol high as the orchestrator, Sol max as Planner & Reviewer, Luna max as task/coder. Very good setup. I'm on pro now and there are weekends when I use half a week's usage but I'll have 5 or 6 sessions going at once for many hours each day.
I don't see how anyone can be using Claude with prices like this, it's pretty incredible what the OpenAI team is doing, w.r.t model quality and pricing.
Prices per 1M tokens Claude Opus 5.5 Claude Opus 5
Cache reads $0.20 $0.50
Input tokens $4 $5
Output tokens $20 $25
Cache writes $5 $6.25
I mainly use Codex/Sol to review my plans drafted by Fable. But beyond that, Astra blows through usage limits too fast to be a daily driver and writes weird code despite what my "house style" is, and Codex is behind Claude Code in terms of critical features like seeing what's going on in subagents.
The parent + subagent workflow has become critical for keeping the reasoning agent (parent) context-lean while also letting me chat to the main agent while work is getting done.
My main process is to use Fable to reason and then spawn Opus subagents, and I get amazing results, and I'm always looking into what the subagents are doing.
- A model (like other GPT ones) that hides thinking traces and thinking summaries, which infuriates me
I've been in the Claude camp for a while, but the way it writes has left me with a a brick for a brain and wanted to see if Astra was as good as they say. Well, I can't know, because in the time it takes for it to actually build anything useful, I've moved to other ideas.
Unbearably, annoyingly slow. I keep thinking I must be doing something wrong.
You're right, it's probably quite unfair of me to say it eats lots of tokens when I am paying double for claude than codex and complaining about tokens.
The rest still stands, though.
But if I've learned anything is that in a 2 months I might have completely turned around, who knows
>I don't see how anyone can be using Claude with prices like this
One potential deciding point is that Claude still has a $200/mo 20x plan, where, since Sept 11, OpenAI does not and has no ETA for the return.
I downgraded my OpenAI plan 2 months ago to the $100/mo, but my usage has gone way up, but now I can no longer upgrade to the $200/mo plan ("This option is temporarily unavailable"). Thankfully I have 2 usage resets available, but I'll probably be switching back to Claude; I was super happy with Astra but I'm burning through tokens and have 4 days before my next reset.
The major difference being the 1M token context window. Once you exceed 272K input tokens, Codex Sol is roughly the same price as Opus; and Astra similar to Fable.
Just because a provider is charging less, doesn't mean their cost went down. This is probably especially true with the big players that are trying to stay competitive.
As someone who has used Claude Code and Codex the prices don't matter in the same way but I found that I burned through my usage way faster on Codex even though I regularly hear that the Codex plans go further. That was not my experience and the intelligence was comparable to what I was getting in Claude.
If these price changes mean that coding plans have effectively more usage then that's great, but Codex is surviving on resets from my own experience using it. I was glad to go back to Claude.
Not that anthropic models are very good at this, but due to the changes in tokenizers and thinking tokens: cost per token is not as helpful anymore as cost / task.
A lot of what I'm doing has pretty expensive build/testing processes between iterations - even on a 40 core machine - so I'm not burning tokens 24/7 like some people may.
I'd guess I'm probably spending >50% of the time running tests & build processes & tooling and the remainder is purely burning tokens.
I also have some internal tooling (that I will hopefully open source soon) that makes LLMs substantially more correct (thus more efficient) - so there's that, too.
and anthropic won't? or any other inference provider? Running your own inference either locally or remotely are probably the only ways to make sure that doesn't happen.
And why is that bad? As your brain gets older, it will not remain so clever, so you'll be grateful for an AI that thinks like you do when it comes to your line of work, failing which the quality of your output could recede like your hairline.
You're building your livelihood/workflows on a set of inputs that you have no idea what they actually cost or how reliable they'll be when the VC cash stops flowing. If you're OK with that, do your thing but it seems a little foolish to me.
Push comes to shove, OpenAI could go out of business tomorrow and I could pick up roughly where I left off for $25k, which is the cost to serve GLM 5.3 Flash on four Nvidia GB10s. Granted, if OpenAI et al go kaput all at the same time, I could probably get a whole lot more compute for a whole lot less money.
In the event of a crash, the investors who put countless billions into this will be still be seeking to maximize their return. Even if it is just pennies on the dollar. Assets (including compute hardware) will be sold, just as they are also sold when any other business fails.
Or maybe a crash doesn't happen. Maybe prices rise to the moon instead and there's nothing we can do to lower them.
Or maybe (just maybe!) a crash never happens and there's never a huge price increase. Prices stay low-ish.
All of these possible outcomes suggest to me that the maximally-sane option that a user can select, today, is to burn it while it lasts. And then, if/when a crash or a massive price increase occurs, just adjust accordingly. (The rest of us will all be in that same boat, too.)
It seems silly to say we have no idea when we actually do, though. We know how much hardware costs, we know how to reliably run a webservice that hits an API hosted on a machine with a GPU, we know how to operate these things at scale outside of OpenAI and Anthropic (not Nvidia). VC money can be patient, Uber's profitable, yeah $1 Uber rides got us hooked and they're running the same playbook. Unfortunately the convenience is worth paying for, so it seems dumb to think we can control the beast or ignore it, or get everyone to agree to hold back.
Is there a world where OpenAI starts charging $2,000/month for what we previously were paying $20 for? What are we going to do? AWS could totally jack up the prices for EC2 instances as well, but we've come to rely on that as well.
GPT would charge more if they could. Both companies need way way more revenue. GPT simply made a calculation that they can earn more money by charging less than their competitors.
They're cutting prices because they want to cannabalize the market for people using models like deepseek via API as well as people paying for anthropic subs.
When they cut prices on luna the first time around they took (literally) millions of users from anthropic.
Any business would charge more if they could. Jevon's paradox would mean that they can make more money by charging less because demand is going to keep growing.
Ya, are LLM's not a great example of Jevon's paradox? I don't think Jevon's needs all else being equal. The paradox being that we should be able to use things less because they are more efficient, when instead they get used more.
Surely, a large part of the increase of the demand in LLMs is in their intelligence, but to hit the demand models needed to be made more efficient, and labs found that more efficient models, still demanded more usage.
These don’t necessarily reflect actual costs, OpenAI is not profitable and nowhere near. They’ve lost their market lead and Sam may feel they need to get it back with any means necessary.
GPT-5.6-Sol, GPT-5.6-Terra, and GPT-5.6-Luna were released in July of 2026.
The first release from the GPT-6 series was GPT-6-Astra. GPT-6-Astra happened on around September 3, 2026, and the previously-mentioned GPT-5.6-* widgets remained available.
Today, September 22, 2026, we now also have GPT-6-Sol and GPT-6-Luna added into the mix.
As I write this, all of the model identifiers I've mentioned are available to select for use within Codex.
Have they solved GPT5.6 SOL's propensity to over-engineer and over-complicate? You'd ask SOL to do something relatively simple, and find four single-use methods, an interface, and a factory-factory.
I actually preferred 5.6-Terra not because it is technically superior (it isn't) but because it had better instincts to NOT do this stuff.
PS - Speaking of better instincts, have they closed the UI-design gap at all? I keep a Claude subscription just because /design produces significantly higher quality UI design/UI feedback/UI refinement than anything I've seen from OpenAI.
I force OpenAI models to use image generation for design, then an iteration loop until it matches the image gen.
This is frustratingly manual and takes many more repetitions compared to Claude (and especially Claude Design) which "just work", but it's a big step change over the default.
The UI design gap is something I’ve noticed as well, in things as simple as ASCII diagrams. Claude has a more human touch. All the diagrams GPT 5.6 generated for me were dressed up lists with too many pipe symbols.
Not the person you're responding to, but I have the same feelings they do and to answer your question for me at least, yes.
IMO 5.6 Sol had this weird dead zone between medium and high where medium under engineered and took short cuts and high over engineered and ignored instructions it didn't agree under the guise of trying being helpful. The whole 5.6 line was the first release from OpenAI where it felt like reasoning level really mattered and was incredibly finicky.
I haven't felt similar issues with GPT 6 though and am very happy with Astra low/med/high as my default choices depending on the task.
In general, I felt like with 5.6 the effort level did less than previous to make the models smarter and more just increased the complexity of the response. I have a half joke theory based only on vibes that OpenAI splitting 5.6 into Sol/Terra/Luna is where the intelligence split happened and so the effort levels were just like "think harder about the decision you already made". So like if the model decided the earth was flat on low effort it'd just say something like "the earth is flat because the horizon is flat". If it was on xhigh reasoning it'd give you a massively complex answer about how the sun reflects light because of the ozone layer and why people flying in planes can see a curve. In both cases though, adding more effort wouldn't get it to realize the earth was round. It just made the answer about it being flat more complex.
To be clear, that theory is not meant to be taken too seriously. It's not based on anything other than vibes. It's just my way of explaining to myself something I'm frustrated about to myself.
Astra 6 was a huge improvement over Sol 5.6 for UI work. I haven't tried Sol 6 yet for it (it's only been a few minutes).
The GPT / Codex models have always been "overengineer" personalities. I prefer that to "I left a pile of race conditions lying around and big gaps in testing" though, which is what I was getting from Opus at times.
But yes both Astra and Sol veer on the side of paranoid. And honestly that's better for team work. For solo work where you just want to yeet something, it can be tiring.
You learn to tame the GPT "personality" on this front by combing over once a week and asking it to find and exterminate pointless tests, clean abstractions etc.
It seems I'm one of the few Terra users since Astra dropped?
When 5.6 dropped I had no weekly limits and I could just drive my work with Sol xhigh and things were great. Once limits were back (and maybe token prices changed iirc) Sol was no longer usable (on Pro or business) unless I was ok with 4 prompts every 5 hours, so I had to switch to Terra medium/high. I've used Luna for some really dumb tasks like moving files, renaming variables and whatever other old-school refactors I've needed.
Then Astra dropped and it just uses so many tokens I've only prompted with it once. Now with GTP-6 Sol/Luna I'm not sure what's being said here but most importantly I'm wondering whether Luna 6 is a good replacement for Terra.
Has any other Terra user tried and knows more or less than answer to this?
Getting to the point where these headlines depress me. I just wish they would stop getting better. I don't know where my career is gonna be in a few years.
+-3 generations of British people lived in roughly the same and in many cases worse conditions (life expectancy dropped during the early industrial revolution, severely in cities). As an average person you would not have been wrong to be against it. It was only in the 1860s-1880s that conditions got better because of bargaining power of the labour class and goodwill of some rich people, two things unlikely to repeat if something like AGI really happens.
Why do they bother creating effort to market all these different models.
All I want to know is how old is the model and how much does it cost. I can figure out which one I want to use based on that, assuming that newer models are always better.
Trying to convince us there is a difference between GPT-6-Sol and GPT-5.6-Terra or whatnot is ludicrous to the point of being insulting, especially when new models come out every week.
Seriously! Though I prefer GPT models to other frontier models, this shit is confusing. They keep changing the names of these models and they often don't communicate anything meaningful about the model itself, especially with these latest iterations. At least with "mini" and "nano" you understood they generally had differing speeds and "reasoning" capability, but what the hell do "Terra", "Sol", and "Astra" really mean? Which one of them is the effective successor to gpt-5.4-mini? It's hard to tell since the only objective information you'll get is token pricing. Is Terra less capable than Luna because it makes me think of dirt and grass? Or is Luna less powerful because the Earth is bigger than the Moon? Apparently that's the real answer. And why do I even have to think about this? And what comes after Astra? Galactica? Or will they start naming the succeeding models after different candy bars? Should I even care since a new model will get farted out mere days after I figured out what differentiated the last one?
What's unclear to me is who OpenAI thinks they're marketing to with this form of branding. These different models don't really mean all that much to the vast majority of people using their products who aren't developers, and developers aren't helped at all by the way they've been naming said models. Are they merely scared that they'll become irrelevant because Anthropic decided to give their models quirky names like "Opus" and "Fable"?
If OpenAI really wants to give their models names, they should name the generation of model and then have the different sub-models named by purpose or capability level. After all, I wouldn't use Mini for a job that Nano could easily do, and I wouldn't use Nano for a job that the full version of GPT-* necessitates. Similarly, I've had to discover exactly how Luna, Terra, and Sol are appropriate for different complexities and task types. OpenAI could help me skip a lot of those steps and just tell me what each model distillation is good for without causing me to look through their pricing page and make educated guesses. After all, shouldn't they not want me to pay attention to how much they're charging me?
All of this makes the days of frontend framework churn seem quaint and actually preferable.
When I use these models in codex, there are two axes for me to control - the model and the reasoning level. I can use Astra, Sol, or Luna. And I can choose between 5 reasoning levels (light, medium, high, extra high, and ultra). What's the difference? As the problems that I want codex to solve get easier, should I turn down the model or the reasoning level? What's the difference between Astra medium and Luna high? So far I've just been leaving it on Astra and then turning the reasoning level up or down based on how hard I think the problem is.
I would describe it as "fidelity" and "verbosity" (or just amount of token generation to complete the task, sometimes that works out to scratch space, or literally how large the "solution" is).
If you have something that needs to be done right, might be a bit complicated, up the model size.
You can see this in the pelicans. Big model pelicans are pretty accurate by default. Up the reasoning and only more so, but with more detail. For Astra, it is 105 lines for low, 250 lines for max reasoning.
Small model pelicans will lack the fidelity of a large model. Bits will be out of place etc. For luna, it's 90 lines for low, 150 lines for xhigh.
Additionally the amount of time taken is increased for the larger models. Luna takes 11 seconds on low, and 1:33 for xhigh. Astra is 33 seconds on low, 4 minutes on max.
And naturally, there is the cost. There's some overlap in functionality between luna xhigh and Astra low in the sense that luna really can do quite a suitable job for some tasks. But there are just some tasks that just don't make sense for Luna, even at high reasoning.
The other thing to remember is that sometimes high fidelity isn't ideal. It can lead to overdesigning. My recommendation is to commit early, commit often, and review everything you do, which we've all been doing since before LLMs right?
I mostly just use frontier models as well. Except for one case: when I let the cache expire (I think 5+ mins of inactivity) I'll switch to one of the cheaper models to summarize and write a handoff note, then pick that up with the better model. Picking up a session whose cache expired with something like 200k tokens with the frontier model reflects really poorly on your usage.
But effort is basically how much extra internal scratchpad to use and how much extra questions to ask and answer before producing a result, Exploring more hypotheses, validating consistencies, calling more tools.
If you’re happy with your token spend on Astra then keep doing what you’re doing. but if you feel the need to conserve tokens, then you can do that by switching to smaller models like Luna when the task is straight forward.
there is no correct answer for that. One is the difference in size/params. The other is the amount of "rounds" of reasoning generating and reviewing what is generated before the model decides it is good.
Cutting prices by 50% as compared to 5.6 prices is exciting. GPT-6 Luna at $0.10/Mio input tokens and $0.50/Mio output is positively insane.
EDIT: this doesn't say anything about availability on either Azure or AWS. I'm assuming it will show up later, but it would be interesting if it didn't.
Per the benchmarks in the post, Luna 6 is at best a couple points superior to Luna 5.6 and (unless I’m reading it wrong) xhigh has actually degraded in quality?
I was hoping for a serious Luna upgrade. It was already cheap enough. This feels more like a price reduction than an upgrade.
That said, if the new Luna is able to handle ultra mode and subagents v2 in codex cli, then at least that’s a win.
Good. Maybe they can use GPT-6 to fix the accessibility of their iOS app. Output shows as text fields to VoiceOver, and the accessibility announcements have backslashes before seemingly every punctuation mark. And then bring accessibility announcements to the Android app so I don't have to make a whole new app just to add that through an accessibility service. Ugh the things I do for accessibility cause I'm blind. On a better note though, AI has done so much for the blind community, from image (and increasingly video) description to mods for video games like Final Fantasy 1 through 6 Pixel remaster, I have a ton to be grateful for.
Luna gets the button wrong, and in the same way Grok/MiMo did. Looking into it more, it's because Luna actually searched my computer for similar builds, found the ones that I did for grok/mimo, and referenced their files. Astra is still the best by a significant margin in my eyes. Far more polish, better page transitions, effects that aren't overcooked and take into account the page. Better contrast.
I just switched from Claude to openAI. I'm surprised at how much easier it is to talk to. Claude always spoke to me with a suspicious side eye as if I was trying to do something naughty. For example I could not get it to help me get an old abandonware game running (sim tower).
yeah I used to love claude! but these days it refuses and responds as if its like a big brother. glad competition exists and for the past few months codex has been significantly better. Even some oss models like glm are good but they dont have enough compute and get capacity constraints
Artificial Analysis is reporting that 6 Luna scores 2 points lower on their coding index than 5.6 Luna, but is 60% cheaper:
> In the Coding Agent Index, Sol improves but Luna regresses: In OpenAI's Codex harness, GPT-6 Sol (max) scores 57 in the Artificial Analysis Coding Agent Index, up 2 points from GPT-5.6 Sol (max), with gains in Terminal-Bench 4.0 (43% vs 37%) and SWE-Atlas-QnA (58% vs 54%). At $2.99 per task it costs ~50% less than GPT-5.6 Sol (max) and sits on the Pareto frontier of Coding Agent Index vs Cost per Task. GPT-6 Luna (max) scores 41, down 2 points from GPT-5.6 Luna (max), with lower scores in SWE-Atlas-QnA (44% vs 49%) and DeepSWE v1.1 (64% vs 66%), at ~60% lower cost per task.
Given that they had to discontinue sales of the 20x Pro plan after the Astra release due to compute constraints, I wonder if 6 Sol & Luna are smaller vs their 5.6 counterparts?
The only two benchmarks shared between the Opus 5.5 and Sol 6 launch seem to be frontier code and automation bench, looks like Sol wins on automation bench (same performance for half the cost) and Opus 5.5 wins on frontier code (2-5% better scores across the board for same cost)
The notable thing is that Luna regressed a bit on coding while dropping 60% in price.That's a fair trade, for high-volume work Luna at that price is basically free, but it does show that newer doesn't always mean better.
Wow, gpt-5.6-Luna was already a fairly unbeatable bargain and now gpt-6-luna is both cheaper and better. And they did a phenomenal job getting gpt-6-sol to max out right where Astra begins; funny how they just so happened to avoid cannibalizing their best model while still being quite cost-competitive near the frontier.
At least it sounds good on paper, the the graphed results do give me pause as it seems the lower cost might come from a slightly nerfed base model combined with more thinking, going by the more erratic scoring curves and the lower no thinking baseline score. I’ll have to try it out but I really hope they haven’t nerfed Luna/Sol to make this price point possible!
I wish there was more transparency on the plus plans usage limits showing actual token usage and prices per model that eats away at remaining usage.
Does anyone know how exactly these price differences for example between sol6 and sol5.6 translate to codex percentages? In theory it seems like for "high" on both it should result in ~3x more usage. If that is actually the case it would be huge! But all we see is % left and % changes while using and we really have no idea when or how those numbers are being calculated or when they change. So there is a 50% price reduction on API but who knows how the hell that translates to whatever price calculation is used on codex.
My initial takeaway is that GPT-6 is mostly a lower cost win, for Luna. GPT-6 Max is an upgrade on intelligence too, but its mostly a cost play (which is great, not complaining).
Overall, I expect for most people think the winner of today was Anthropic. I personally am preferring Opus 5.5 at medium over GPT-6 Sol Max, in very very early tests. Similar price range, more capability.
But competiton is great, these are solid releases by OpenAI today.
It's asking a lot to trust they can or will maintain this new pricing. In any case it's exciting to think this might lead to further price cuts in the highly competent and competitive Chinese clones. I'm still using ChatGPT for interactive queries, but at this point pretty much only because of its familiar UI
SpaceX is a different beast with extremely high friction to enter its market, a massive technology lead, and well developed preferential high level relationships with just about every country worth worrying about.
OpenAI/Anthropic meanwhile feel a bit like they're hoping to sell iPhones in a market about to be flooded by $20 flip phones, with almost no channel of their own to do it. And for whatever mad reason OpenAI are now signalling they will attempt to compete on price with flip phones despite their cost of labour, energy, and just about everything else being far higher
Wasn't SpaceX's insane valuation largely based off of Grok, because their rocket and satellite businesses could never be valued at over a trillion $$?
I don't see your metaphor to iphones and flip phones. This new Luna model is cheaper than deepseek 4.1 flash, except for cache reads. OpenAI having to compete with China is a much larger economic-political issue that is far larger than just our AI labs.
Well, they do rug pull constantly. This week and last leading up to this the cost to use Codex was overwhelmingly perceived as terrible. People running out of usage all over the place. Reddit full of people crying. I noticed it myself.
Then they do a new model launch, issue quota resets all around, and it's a party for 2-3 weeks before things return to normal.
Since I spent my morning fixing a bug in my OpenAI API proxy that completely broke prompt caching and caused my usage limits to burn like kindling, really happy to see some of their new cache tooling:
I’m still finding context is king, even with the best models.
For example, I had Fable review Astra’s output yesterday, and it found some issues and fixed them. Passing the fixes back, Astra then uncovered additional issues with Fable’s fixes (and yes, this will go on ad infinitum if you let it, but these were “real” issues).
It seems the big story here is the reduced Luna pricing. It’s a fantastic model that can handle most automation needs (though I still use the big models for day-to-day development).
Not a source but a comparison with a weaker non-SOTA model:
Nvidia's top AI chip Rubin sells in 72-GPU racks for about $3.5–7.8M. A rack running Xiaomi's MiMo V2.6 Pro generates roughly 150–300B tokens a day, worth about $130–260k at Xiaomi's API price. That's a payback of the infrastructure in a few weeks in theory. After a few weeks or a month, the only cost is electricity, and whatever they make after that is pure profit.
OpenAI and Anthropic are practically scamming people with the token prices.
Nah, Luna uses was more tokens and fills the context up way to fast. Terra is in the sweet spot where if feels like Opus 4.6. Competent but not too smart. It also lets you have longer sessions (back and forth) without filling the context too fast.
Except for the censorship. We use it for massive data crunching, and roughly 5-8% (depending on the day) gets censored and doesn't get a response. We switched to Mimo 2.6, which is relatively better. For censored stuff, we use Sonnet and OpenAI Nano models.
people are forgetting how many copilot licenses are sold coupled with gpt models, adoption is pretty low and they are making ton of money on that, "allocation of unused tokens"
The hard part for me is choosing the model and effort, that's why I always resort to Astra xhigh, but then it ends up consuming tokens so fast.
How do you decide what to pick? I mean, I do Platform work on a large monorepo with many different interconnected services, and so I always want the implementation to be "correct".
hmm, it somehow continued the trend of being basically the same score on https://artificialanalysis.ai/ as the 5.6 variants.
I kind of hated Astra for it's poor instruction following and stopping all the time plus bad code quality. It somehow feels a bit like some of the popular open models but with a lot more knowledge or peek capability. But it doesn't reach peek that often
It wasn't really used enough and it sat in an awkward middle space between luna and sol where either luna high/xhigh or sol med were better cost/perf wise
Your employer should reconsider. Sol high is cheaper than Terra max and smarter, when measured per task. ie even if tokens are more expensive Sol can often do a job with fewer tokens.
It probably didn't see that much use, as it struggled to find a niche. If you wanted intelligence tasks, Sol was cheap enough and much smarter. If you wanted performance and cost-effectiveness, Luna was significantly better value while being only a little less intelligent.
Terra ended up just being an awkward middle ground that was not particularly suited for any workload.
I disagree. After a bit of experimenting, I actually found Terra to be a very good workhorse model on none-to-medium reasoning, and I actually quite prefer its code to Sol's in many cases. It has less of a complexity to over-complicate things. Where Sol would have a sea of try/except and recoveries for situations that are structurally impossible, Terra would just write nice, sequential code.
Maybe for one-shotting large things Sol is better, but for prod code where I decompose into smaller tasks and read all the code I favored Terra.
Sol 5.6 was still king for architecture/research in my workflow, though.
When using max effort, I run into context compaction quite a lot. I haven't seen any increase in context window size at all over the past half year (stuck at 1M) for all the flagship frontier models.
Have the frontier labs stopped trying to increase context window size?
I've found Astra to be horrible at making orchestration decisions. I will be trying to use Sol for both. Fable is very good at it though. Worst part of my week is when I hit my Fable usage limit and have to switch to Astra.
1. you can have luna clean up after itself and improve code 2. you might be doing something like video-editing, cad modeling, artistic direction, pcb routing, etc. that need to run a long time to "converge"
Has there ever been an instance in history when this strategy worked?
Plato argued that writing things down will make your memory worse, and less skilled as a debater (kind of true!)
How are the Luddites doing at textiles?
I remember the arguments that using 'high level languages' like C and Pascal will make you not understand machine specific details (kind of true!)
I respect that you want to learn how things are done, that is a great trait. But once you learn how its done, you should use the tools to free up cognitive load for more difficult tasks.
AI needs to get rid of model versioning and model effort combinations. It's like selling an automatic transmission but still asking you to choose the gear, then after the trip telling you how much fuel you burned.
A good router on the user-facing end would be nice, but I’d rather be able to get a feel for which car I’m driving and pick one depending on the task, than have to hope the rental agency knows what I need.
What a day! I couldn't really use the last Luna for much (wasn't smart enough) or Astra (too expensive). So this release is really exciting. I can probably use Sol 6 as much as I want in the week, which as great.
If I'm understanding correctly now when you want to use Codex to do a given task you need to decide between:
GPT-6 Astra (low medium high xhigh max ultra)
GPT-6 Sol (low medium high xhigh max ultra)
GPT-6 Luna (low medium high xhigh max ultra)
And that's not even counting the GPT-5.x models:
GPT-5.6 Sol (low medium high xhigh max ultra)
GPT-5.6 Luna (low medium high xhigh max ultra)
GPT-5.6 Terra (low medium high xhigh max ultra)
GPT-5.5 (low medium high xhigh max ultra)
And then there's a fast mode toggle for all of it, too.
Not exactly a low-friction user experience!
Like are you supposed to just somehow intuit, “ah yeah, this task is definitely a GPT-6 Sol Medium task,” or something?
Is this just second nature for OpenAI employees? How are end users supposed to know how to optimally choose a model for a given task? Am I missing something completely here?
It’s confusing indeed, but I like having many options, particularly considering that pricing can be wildly different depending on the model.
Maybe OpenAI can offer an "auto" mode for Codex on the subscriptions, while leaving the possibility of users manually overriding whatever model the router chooses. To me that would be the best of both worlds. The problem is building a competent model router.
Looking at their own charts it seems like it's only small incremental improvement over 5.6 Sol, but with a massive cost reduction. And the better writing/communication style that Astra had.
Am I the only one that doesn't really feel a difference in performance from model to model?
From around GPT 4 results got "Good enough"...I generally try to explain what problem I'm trying to solve, set limitations and boundaries, tell it to ask me questions, have it write up a plan with steps then we take one step at a time.
These new models are starting to feel like iPhone releases where the improvements / feature set feels incremental.
Is this comment about astroturfing or a decline in comment quality? To be honest, I was one of those early commenters, and I was just genuinely shocked at the price drop. I am also excited to try Opus 5.5!
Fundamentally I feel like coders just doesn't even need to be that smart anymore given AI assistance. This place ten years ago used to be filled with some of the most interesting comments/takes around for that reason.
Ya all these articles lately about how everyone is sick of reading AI prose, and interacting with models in general. Tons of new model optimizations and workflow optimizations or whatever. I'm not really aware of any idea or product aimed at making the internet usable, and making it somewhat resistant to the generated noise. I think HN is a bit better than reddit for this type of example for floods of comments, first movers on reddit REALLY rise to the top and stay there.
I remember how a few months ago Dang was criticizing people for making comments like this. Guess he just realized how stupid that was and stopped bothering eventually.
Are you trying to imply that nothing OpenAI can release would justify that response and therefore the people must be bots?
Asking cuz I don't think I'm a bot [pats self], I legitimately prefer the GPT models to Anthropic's, don't like Anthropic's customer service/reliability story at all, and I welcome a massive price reduction. Seems like something I should be happy to get.
If you'd told me I'd be typing this a year ago I'd be skeptical though.
But the reason people say "Claude can't compete" is because Claude Opus has been going downhill since 4.7, and many have found Opus 5 intolerable. Fable is much better, but also much more expensive than OpenAI's offerings.
OpenAI is promising "the Sun, the Moon, and the Stars". The spirit of P.T. Barnum is doubtless looking on with jaw dropped at what is beyond doubt one of the greatest demonstrations of chutzpah, by some of the greatest hucksters, in the history of the human race.
GPT-6 has been fantastic to use. I see Opus 5.5 today but honestly it's been such a rough year with Anthropic, and OpenAI's models are so far ahead, it's tough to consider moving back. I also think OpenAI's desktop app is significantly more polished than Claude CoWork.
This is all just running in circles. The models are not obviously better. The pricing fluctuates or offset by some other less-obvious metrics (availability/speed/tokens per task/dumbing down). Everyone reports different outcomes in their usage because it's all so context and user dependent. Sometimes models do some things better but become so annoying and obtuse in their other doings that it's just not worth it (like Opus with the insane code comments and Astra with its over-the-top, everything-is-a-sales-pitch style). It feels like the AI gods just turn the knobs on things like compute to get the results they want to align with the IPO to make headlines.
This is the most blatantly astroturfed thread I have ever seen on Hacker News.
My previous comment--which suggested astroturfing--was the highest upvoted comment here until it got flagged. Which implies to me that atleast the other remaining humans on this forum see it as well.
26 minutes, 89 comments, upvoted instantly to the top, posted within one hour of the Opus 5.5 announcement. You tell me.
There's no evidence of astroturfing. The comments you’re referring to are from accounts with established history and in different locations, without any evidence of being linked to OpenAI. They just seem excited about the models and the pricing.
On the other hand, you have previously written: I'll gladly admit I think what these companies are doing is unethical, and I'm sure that biases my thinking toward skepticism. [1]
You have now posted accusations/assumptions of astroturfing and manipulation at least 15 times, without ever providing any evidence. This is in breach of the guidelines, because comments like this poison discussions far more than the comments they're complaining about.
We – of course – want all comments and posts on HN to be authentic. HN is only a place where anyone wants to participate because since the beginning, we've had software mechanisms and moderation practices that detect and weed out inauthentic commenting and voting. We're identifying and dealing with it every day, continually improving the software to detect and remove it. Most of that happens quietly and efficiently in the background without anyone having to see it. When users see evidence of manipulation and report it to us via email, we happily and thoroughly investigate it.
Most of the time, what we find is simply that people are authentically excited and passionate about the topic, which is what is happening here. I understand it can be hard to accept that if you're skeptical about the topic.
It's fine to be skeptical about the topic and you're welcome to express your skeptical views on the topic. People do that every day on HN, about AI-related topics and countless others. Healthy debate is what we're here for.
But you can't keep poisoning HN, by (1) continually posting these unfounded claims, then (2) when users and moderators simply uphold the guidelines, staging a protest by demanding your account be deleted. This is not what people do when they care about a forum's health.
Hopefully it is clear from my other comments that I do try to provide value too.
> You have now posted accusations/assumptions of astroturfing and manipulation at least 15 times, without ever providing any evidence.
I tried to point out the upvote speed and age-to-comment ratio for this thread look anomalous to me, and that it being posted within the Opus 5.5 release hour was further reason for skepticism. Circumstantial, sure, but I see very little ways to gather hard evidence of astroturfing without being a mod.
> when users and moderators simply uphold the guidelines, staging a protest by demanding your account be deleted.
You're right this was a little dramatic. I think it is just annoying though when two times that I have posted about astroturfing, it has been the most upvoted comment only to get flagged. I guess this is a self-fulfilling prophecy though, as you are right that other human commenters are abound and tend to flag people complaining about astroturfing.
Anyway thanks for the reply. If your read of my account is that I'm more-often-than-not a bad actor, then I will stop commenting here. Seems like it is for the best. :)
Like with dang in his interaction with you a week or two ago, it's pleasantly surprising and welcome that you respond so cordially to our replies.
> Hopefully it is clear from my other comments that I do try to provide value too.
I agree that you provide value in plenty of your comments, which is why we don't just want to ban/lose you.
> I have posted about astroturfing, it has been the most upvoted comment only to get flagged
People love a conspiracy theory, and on a site like HN that has many people looking at it at once, it's easy to get a large number of upvotes in a short amount of time if enough people find it exciting. We often see off-topic, titillating one-liners or ragebaity comments at the top of threads, and we always have to downweight them to keep the discussion on-topic and healthy.
> Anyway thanks for the reply. If your read of my account is that I'm more-often-than-not a bad actor, then I will stop commenting here. Seems like it is for the best. :)
It seems like you're well intentioned. You have your concerns about A.I., as many do and that's fine. You're still very welcome here. Just please try to believe that many or most of the people who are enthusiastic about A.I. are as sincere in their positivity as you are in your concern.
Do you think the "you can't mention astroturfing" rule is really serving HN these days? Do you think this thread hasn't been manipulated?
I respect you for replying here though, and yes I get that HN forum standards would suggest flagging my previous comment. But it is just sad to see a place used to be so vibrant get manipulated because of how much weight it holds for us in the industry.
And yea sure, I could go and flag all the bots and message Dang. But probably time to stop shouting into the void. :)
Yes, this thread is not being manipulated. People a) being excited about something and b) it being in favor of a certain company is not sufficient evidence of astroturfing.
I've seen your articles in the past and I thought they were great. So I respect your thinking, and have no intention of being combative.
But do you really think people were so excited about cheaper versions of Astra that they were just waiting around to comment the instant this was posted? More than two comments per minute? All the initial comments were really similar too: brief one liners celebrating the cheap prices.
I think AI right now is a sort of Rorschach test. What it is clearly revealing to me is that I don't trust organizations with enormous financial incentives to not manipulate public opinion. So I see bots everywhere. :)
>Do you think the "you can't mention astroturfing" rule is really serving HN these days?
yes, i think so.
because, unfortunately, complaining about bots (or astroturfing, or whatever) doesn't stop them. so we end up with threads that have both the potential bot/astroturfing/whatever activity and complaints, which further drowns out any interesting comments.
Do you think all the comments in here are positive about the model? Because they aren't. In no way does it seem astroturfed. And yeah, pretty boring to read that kind of comment every time.
Engagement is generally more important than purely positive sentiment.
Anyway these comments were made when this thread was in an earlier state. I agree that it has gone on to be more "organic" looking. That doesn't exclude it initially being manipulated to the top, in my mind, but certainly they aren't carpet-bombing with only booster comments.
More than two comments per minute right when it was posted, all simple one-liners celebrating the price decrease.
I would agree now that the thread has recovered to a more interesting state, but how it first looked--combined with it being posted right after Opus 5.5 announcement--look questionable to me.
You have to accept that people are different than you.
I see long massive pro apple threads. I don't get it at all. As in; literally don't understand what Apple is good for. But I have friends, family irl who love apple so I know the sentiment exists. I therefore accept that many HN users are similar.
Many people really truly are happy to see another model drop and are excited about progress etc etc. Surely you've met such ppl in real life. Well, they're here too (I'm one of them fwiw)
You know I think I have accepted this after a few decades on this planet, but probably can't hurt to be reminded. :)
I wouldn't argue there aren't real humans excited for this drop. It was just all the circumstances around it--the comment speed, the upvote speed, the initial uniformity of what people were saying.
Clearly this isn't the first time. Remember criticizing model releases will get you flagged here.
This is why HN has been on the down hill in quality and those that care to highlight that are being punished, while the astro-turfing, gaslighting and Show HN self-promotion slop continues.
Here's GPT-6 Luna pelicans: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
And GPT-6 Sol: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
Scroll to the bottom for the GPT-6 Sol max one: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
For comparison, here are the pelicans I got for GPT-6 Astra: https://tools.simonwillison.net/markdown-svg-renderer?url=ht... - I still like the Astra Max one best.
Here's a comparison grid showing all of the GPT-6 and GPT-5.6 pelicans at all effort levels: https://static.simonwillison.net/static/2026/gpt-6-and-5.6.h...
The grid is actually really interesting, because it shows that the 5.6 family default to brighter colors than the 6 family.
Cached Read: ~6,500M
Input: ~150M
Output: ~20M
Approx $40 worth of usage across DeepSeek V4 Flash + MuseSpark Contributor 1.3. And a bit of both the GLM models. This is covered in a $10 subscription.
If I were to use Luna's API pricing:
$0.02 x 6,500 = $130
$0.20 x 150 = $30
$1.20 x 20 = $24
So $184. And this is assuming smaller coding sessions (<272K) beyond which Luna pricing doubles.
--
Cost wise, these models are nice for small stuff. Translations etc. Any model that does not provide multiple Mtoks of cached reads per cent is not very useful to me for coding workflows.
Snark aside, Luna 5.6 was (is) an incredible game-changer.
perhaps it then does mean - squeeze as much as you can get off this actual free usage.
When I ask for an explanation it adds the right amount of detail. Of course, some of the material is new to me so subtle errors are hard to spot. But at least I’ve caught Terra and Sol on inconsistent messaging.
Also I’ve found 3.8 flash to circle back to root issues even at the conceptual level like problem fit and conceptual solution direction or architecture when I wasn’t achieving my goals. It flat out said I was attempting to use the wrong tool. Whereas Sol and Astra kept rabbit holing and looking for tiny implementation errors. Even after prompting them specifically to look at it broader.
By raising it from investors.
https://en.wikipedia.org/wiki/Pareto_front
https://en.wikipedia.org/wiki/Pareto_principle
I assume it's a subsidy to get more training data.
EDIT: Okay downvoters, what's your take on why they're giving away Luna for so cheap?
GPT-6 Luna High: https://threejseval.com/models/gpt-6-luna-high
GPT-6 Sol coming soon on the site.
You can compare any other model on the same prompt. Gallery unlocks after 4 votes: https://threejseval.com
I had all the tabs open individually and harder to scan which model is which... otherwise keep up the great work! I like the grid view a lot. (Also the pages have no OG images set, which impacts what the link looks like shared)...
OG images will require me to move away from publishing in a Gist and linking to from a JavaScript page that loads the Gist. Probably worthwhile though.
Everyone said tokens were too expensive but these are getting close to free while still having fantastic performance.
1 - https://bench.killswitch-lang.org
Is it? It was already too cheap to meter for me. Luna 6 is actually worse on some benchmarks than 5.6. I’d have loved improved performance for 2x the price than ~equal performance for 0.5x the price.
I expected a Fable 5 -> Opus 5 situation, where GPT 6 Sol would perform on par with GPT 6 Astra.
Instead it's more like a price cut on GPT 5.6 Sol, and I'll have to stick with Astra for my work.
The only thing I can hope for is that more users switching to the GPT 6 Sol model frees capacity, allowing OpenAI to hand out some usage resets.
They put both legs on the same side of the bike.
Even Astra max which actually put one leg on each side of the bike still somehow messed it up because when it added the bike chain, it put the left leg between the bike chain and the frame.
Is what I'm getting on the top two links.
Half the price when it launched, or after the price dropped by 75%?
> GPT‑6 Luna vs. GPT‑5.6 Luna | $0.20 → $0.10 | $1.20 → $0.50 | 50% cheaper
I can read it as follows (below), meaning that GPT-5.6 is 50% cheaper.
- GPT-6 = $0.20
- GPT-5.6 = $0.10
good god
I'm so tired of looking at benchmarks. I always look fwd to the pelicans.
And Astra medium seems to yield similar or better quality for the same price as Sol 6 xhigh.
1/ Usage limits: downstream of input/output cost, but resets and obscure windows and odd 20x plan / 5x plan != 4x usage math throw a wrench into it. Winner right now is Codex by a mile, especially when you factor in the fact that ChatGPT usage (even 6 Astra Pro) is essentially unmetered on the 20x plan. Always a bummer when asking if I should see a doctor about a rash means I can't code as much. It's also is a godsend if you use an MCP like oracle to automate the process of calling the Pro model on particularly tough problems, giving better planning results or deeper code analysis without burning usage.
2/ Context window in the harness. Claude Code wins on this. There used to be a toml file workaround for Codex to extend the GPT context window to 1m, but this stopped working on the plans and only on per-token billing (ETA: noname120 pointed out this is no longer the case and it can be enabled again [1]). 252k is just not enough. Codex's compaction is very good, fwiw, but it happens so frequently that even a model as powerful as Astra sometimes loses the plot on long-running tasks.
3/ Ability to use the plan outside of the official harness. Codex wins. Anthropic does shit like bills requests as extra usage if it sees a hermes.md in a commit.
I've subscription hopped a bunch, and at times I've had both, but I keep coming back to Codex because it wins on 2/3.
ETA: apparently I haven't been Keeping Up With the Altmans and new 20x signups have been disabled for a few weeks. I am grandfathered in, which makes the comparison above pretty much moot.
[1]https://news.ycombinator.com/item?id=49806060
This hasn't been the case since around July. If you measure usage in raw api costs, Anthropic is actually giving more on $200 than OpenAI now. This includes resets. Usage allocation difference would be humiliating for codex subs were it not for resets. But fixing usage limits with resets is ugly, and they're not good for your mental well-being.
> Context window in the harness
Codex now allows 1M for subs with config params. But generally speaking, you shouldn't really be using 1M context. If you accidentally send a request with say, ~700k context already accumulated in a session which is outside cache TTL, you're paying full cost of these 700k tokens.
> I've subscription hopped a bunch
OpenAI actually has a new strategy to prevent subscription hopping after their 2-3 month-long marketing push to get claude-folks to switch over:
you can't buy a $200 sub anymore. So if you cancel, you won't be able to get back in. Hostage situation, essentially.
EDIT: re: usage limits, oh-my-pi maintainer has been tracking this - https://nitter.xitter.cc/_can1357/status/2090075496948060372
When i swapped between a 200k Fable context into an Astra model (i was out of fable) the token usage in that context dropped to 150k or something.
Either there was a bug somewhere, or the same text got cut up very differently between providers.
I think I’m gonna move back to a Claude plan. I could barely hit the $200 limit if I went non-stop on programming tasks.
When they started the aggressive campaign, entire X (including myself, sadly) was full of posts about how "unlimited" codex usage is even on a $20 plan. Sam Altman was posting something in line of "we love our users, unlike Anthropic". Got my network to get codex subs because of the value compared to claude.
Then they gradually reduced the limits to the point where even $200 plan only lasts you just 1-2 days and $20 is basically unusable, then the hostage thing.
Do you have /fast enabled by any chance?
I was considering the $100 plan, but I hit the 5hr limit in an hour. So even with the $100 plan I figured I cant go non-stop on a single agent running Sol Medium
I am considering the plan myself. I just don’t know if I want to fork out $100 per month for something I will make $0 off of.
Opus is at least actually usable even on the small plan. The main downside is its insane writing style, but 5.5 seems to address that somewhat. Otherwise, you can just use your $20 OpenAI plan to have Luna de-slop Opus' prose, which seems to work fine.
...but the few times I've tried to use codex for a moderately difficult task it burned through its limit extremely quickly.
But even if we leave that aside, OpenAI models are also much more eager than Anthropic, which are on the lazier side. Left unsupervised, Sol/Astra will attempt to build a sha256 verified rocket ship if you ask them to fix a race condition in your to-do list app. Anthropic models will do what you asked for, maybe even forget to implement parts of that ask, but they won't generally throw a slop granade at you.
I can leave Fable orchestrator unsupervised for ~2h. Leaving Sol/Astra unsupervised for ~2h means the next user turn will contain a message: "what are you doing and why?".
I haven't been tracking, but this roughly matches my experience with codex 20x and claude 20x subs. Claude subscription now lasts me 3-3.5 days on average. Codex is 2-2.5 days. This is work on same projects, with similarly sized tasks.
To make matters worse, I've merged a lot more code produced by fable than sol/astra.
Are you sure?
https://x.com/thsottiaux/status/2098113585683808624
Now, if they disabled it yet again, that's another story. But that tweet is not evidence of that.
Though with the price of GPT-6 Luna, the temptation to switch to pay-per-token grows.
It's been disabled for some time now though otherwise, I check about once a day myself and keep and eye out on social media.
Annoying since I was about to upgrade back to the $200 plan after downgrading to the $100 plan due to being on leave and not needing as much usage the month prior. Doh.
Just checked my toy chatgpt account that only ever had a $20 sub. $200 plan still shows "The 20X plan is temporarily unavailable for purchase".
While you're understandably not including the values of the $20 standard plans on both, I find the generosity of then token limits on ChatGPT plus vs Claude Pro (it's a huge difference) to be good representation of their respective attitudes towards the average user. You literally cannot use Claude pro to build real software, unless you're extremely frugal with your prompts and don't try anything even a little ambitious.
Also, Anthropic has zero models comparable to Luna.
They're both pretty horrible, but I find it difficult to find arguments for why Anthropic is worse than OpenAI, other than their doomtrolling. Which, in the grand scheme of things, doesn't even register.
Edit: forgot about the SpaceX thing.
Anthropic is trying to kill open models way harder
That alone is reason enough. Also, I don't think either of them are horrible. That's honestly a ridiculous take considering how much people in here love their models, and how much they've advanced the industry forward.
I often have the urge to design my own harness too (once I have more time). But even with the current mainstream harnesses out there, there's just to many hurdles if you wanted to mainly stick with anthropic models and need the subsidized pricing (from a sub).
That's a non-sequitur.
"Nestle is a great company, considering how much people love their chocolate."
Yeah I don't think the handling of copyrighted training data was correct, but I can't pretend I know what the correct solution to that issue is.
Speaking of OpenAI specifically, they don't price gouge people, they aren't aggressively anti-competitive, they're not nearly the perpetual hypocrisy machine that Anthropic is (which is one thing I actually really dislike).
Regarding Nestle, it's pretty obvious that the sentiment towards them is a lot more negative and they aren't universally loved by any group of people. Processed foods are by and large garbage nobody needs. Their use of forced labor is denounced by just about everyone. What have OpenAI/Anthropic done that's even similar in scope to the forced labor / modern slavery that people hate Nestle for.
If you had a company that genuinely helped hundreds of millions of people worldwide become more productive and more satisfied with their tools, and the overall sentiment towards your products within the industry is positive, then what argument would there be that your company is "horrible"? At least give some decent counter arguments.
You mean aside from "the largest theft of labor in human history"[1]?
[1] https://www.nytimes.com/2026/09/17/technology/microsoft-open...
Anthropic does all that but they're also populated by many people who believe they are building God and that they must build their god first in their own image so that it can take control of humanity and protect us from any competing god which is not built in their image. Their position is inherently paternalistic and authoritarian, and they consider suppression of competition not just important to the bottom line but to life in the universe. Under the doomer ethos there is no evil too great to rationalize.
There are plenty of wrongs done in the name of profit, but capitalists have nothing on zealots in terms of causing serious harm. Profit motives can be directed by influencing incentives, but zealotry is frequently terminal.
That isn't to say that there isn't some overlap-- the cultists have infected both organizations. But OpenAI has pretty consistently only given lip service to AI doom to the extent that it improves the bottom line, while (mis)Anthropic was founded specifically because OpenAI wasn't mentally ill enough.
Anthropic has great products, but it's not meaningfully better to 99% of devs that I'd rather support the company that doesn't constantly act in opposition to optimism and to the vibe I'd prefer for a 100 billion dollar (or however ridiculous amount they're worth now) tech company embraces.
AI doomerism is a genuine waste of time if you aren't actively pushing towards a better AI industry for everyone, not just the groups in full ideological alignment to your personal leanings.
Interestingly I would have drawn the exact opposite conclusion looking at my Claude and codex usage.
I can't get anything sustained out of codex in chatgpt plus, while I have been using Claude pro extensively and put on a lot of experimental task and features.
I ran into codex exhausting a 5h window on code review in minutes (like 3minutes) multiple times, while I could get Claude to implement 2~3 medium sized features with the same usage consumption.
(I also really dislike the usage resets in codex, they always make me feel like I use them wrong because I often just want to reset the 5h window, but they can only do both at once...)
> Also, OpenAI is just a company I'd rather support than Anthropic.
Sure, he's free to say whatever especially considering the amount of revenue he's creating, but it's just an altitude that I prefer not to see.
And Dario's "AI will kill us all" is the same as Sam's "AI will discover ALL science and we'll be building Dyson spheres".
Different flavors of the same BS.
"You literally cannot use Claude pro to build real software, unless you're extremely frugal with your prompts and don't try anything even a little ambitious" - that's way past ridiculous. Even just using Fable most of the time, working on several ambitious projects, I have a hard time hitting the limit with a Max plan.
I'm quite puzzled about why Anthropic is so hellbent on blocking other coding agents. It's not like Claude Code has any secret sauce, right? And doesn't Anthropic make monkey off API usage, and their magic is on the model side anyway?
But to be fair, they don't really enforce the harness rule that much anymore. I guess if your harness doesn't do a lot of weird things like a lot of cache misses, or triggers some distillation attacks, or some broader Chinese fingerprints, they're tongue-in-cheek okay with you using a third party harness.
oh-my-pi supports it natively (again, still a ToS violation), by impersonating claude code's fingerprints.
I have been using oh-my-pi with 3 claude subs for the past few months without any issues. Even native server-side OAI/ANT compaction works out of the box.
As far as I know Codex (at least the GUI) can automatically call the ChatGPT Chat models (including Astra 6 Pro), you just need to @ a ChatGPT Chat conversation from within Codex and tell it when to use it.
> There used to be a toml file workaround for Codex to extend the GPT context window to 1m, but this stopped working on the plans and only on per-token billing
Not true, it works again[1]. I confirm that it works both on 5.6 Sol and Astra 6, possibly other models too.
[1] https://x.com/thsottiaux/status/2089082893804896524
And re: the toml workaround, AWESOME! I appreciate you pointing these two things out, this is my highest-ROI HN comment thus far.
Having limitless webUI ChatGPT usage is much better user experience, though. I'll give them that.
(edit: Sol-6 is half the price, so maybe the usage limits are going to be way better.)
Meanwhile I just burned ~20% of my weekly quota with Astra making one config file for a service.
I appreciate and follow Matt Pocock's advice: avoid autocompaction. Compaction is lossy, which is ok when you're managing it at phase boundaries, but autocompact is lossy at the most inopportune times, firing mid-task and leading to agents going off the rails.
My conversations compact hundreds of times. By the time it has done a dozen or so compactions, it fully understands the work I want it to do (and how). It's almost like having a fine-tuned Astra model.
10/10, would recommend.
The GPT was about 1B on two projects on 300$ worth of plans all on Astra and I capped out on usage.
Anthropic caching must be better because the cache rates are better on Claude models.
Compare that to Claude and I can run multiple agents on Opus almost indefinitely. YMMV of course but I was shocked at how quickly I burned through Codex usage.
On the context window, I feel so cramped on Codex, compacting happening every time I turn around is annoying. I didn't realize how much I enjoyed the Claude context window size.
Makes me think they picked Codex, stopped trying Claude, and just hang on to outdated beliefs about the value they're receiving.
I am curious how the 5x plans differ between both providers.
Opposite in my experience. I need to limit codex to 500k on medium/low, still run out in 2-3 days with 1 CLI window. CC gives me 4-5 medium/high days with 2-3 CLI windows, and Opus is still great for other regular dumb engineering/refactoring.
On the other hand my head starts to hurt if I read Opus for too long, hopefully they fixed it with 5.5.
Using Agentsview (which might have it's own issues) I was getting ~$200 of API usage in my 1 week Codex window (paid $100) vs ~$5,000 of API usage in 1 week for Claude (paid $200).
And ya i can go over that 240k limit, I still very seldom do, and try to treat it as the actual limit. I'm surprised to see so many people still talking about compaction to complete long running tasks, i think the bulk of the work should be somewhat frontloaded into a plan that is split off into subplans, then you can kinda open up a few options, one session with subagents for the subplans of the main plan, or just handoff prompts about progress against the main plan/relevant subplan. I just never trust the blackbox that is compaction, I feel its a recipe for disaster/context poison.
I’m currently on the 5x plan and burned through 5% today on a difficult task in 15 minutes so I doubt that. If you got the wrong kind of tasks that you work on, it can go fast.
LMAO, I wish this were true, I hit limits (and the "we are disabling access to protect your data" warnings) all the time, or have chats just...fuck off and get into weird/invalid states (interrupted chats, chats that are spinning and stuck, returning "/mnt/" paths instead of images/md files, file links being returned with no file backing them, image classifier firing...and then returning the image anyway (though now I know that GPT-Image-X really really wants to generate NSFW even when that isn't the request)).
Though I am probably an outlier, I have both 20x Claude/ChatGPT plans and max both out every week, so... (in my defense I am a hobbyist and this is out-of-pocket)
Is this not the default anymore? I am on the (now closed) 20x plan.
That isn't a valid comparison, since Codex 20x is closed. So we should be comparing Claud 20x to Codex 5x + credits.
Also in Codex, even though you can increase the context window to 1m so its on par with Claude, exceeding the default is billed at 2x.
It wants things beyond what the mortals (us) know to reach for. It's not good at explaining itself, it doesn't show it's thinking. It's often not wrong. But the no compromises attitude can be unbearable to deal with. Especially given how little it cares about telling us.
The other explanation is just as part of ‘token efficiency’
Otherwise I'm using 5.6 Sol for actual plan execution and review..
Astra was/is superior for planning type tasks. It was capable of doing seemingly magic things with rather vague/lazy instructions ("I need to be able to test this on Windows, maybe a qemu VM or something? Shrug." ... 1 hour later "yeah i built you a whole qemu + eval windows image + harness of powershell scripts + shell scripts to retrieve & verify harness.").
And for UI work -- which is not something I do a lot of but do here and there -- it was clearly superior to 5.6 Sol.
But it also feels sloppier? Somehow. And too expensive to use.
We'll see how Sol 6 is.
Terra had the "workhorse" quality where it could do these changes in bulk and follow directions without being too 'smart' (but sloppy) as you described. Luna was a bit too dumb and would make sloppy mistakes; I see that more as a "run these tests and format the results" sort of model. Maybe 6 Luna will be better.
I also just reread your comment and realized the naming convention is still extremely confusing with respect to ordering of [Family]x[Model]x[Number].
When people spend their days interacting with machines that pretend to be human, they may then start treating real humans like machines.
If I leave Astra overnight, I'll wake up with three new different projects, each of them 20% done and having nothing to do with my original goal.
Like you said, it felt very natural to work with. Opus 5 is way too slow and verbose for me, I find I get distracted and annoyed with it.
Opus 5.5 seems a LOT closer so far to what I liked about 5.6 Sol but we'll see
Highlight and lowlight of my week was successfully convincing the OpenAI support chat robot to give me a refund for the month for my issues with 6 chewing threw my usage with no output.
- For general chat and web search, occasional image editing, small coding work, document review etc. ChatGPT Plus is basically limitless and “just works” since 5.6. I’ve yet to give it some task it cannot do.
- When given sensible instructions, it hardly annoys with weird phrasing, glazing, or annoying constructs.
- The apps are very good (ignoring the initially terrible Codex app)
It’s easily my best spent $23 a month.
Still needs a LOT of work IMO.
Input
Output
Price reduction
GPT‑6 Sol vs. GPT‑5.6 Sol
$4 → $2
$20 → $10
50% cheaper
GPT‑6 Luna vs. GPT‑5.6 Luna
$0.20 → $0.10
$1.20 → $0.50
50% cheaper
The parent + subagent workflow has become critical for keeping the reasoning agent (parent) context-lean while also letting me chat to the main agent while work is getting done.
My main process is to use Fable to reason and then spawn Opus subagents, and I get amazing results, and I'm always looking into what the subagents are doing.
- Unbearably slow
- A token eating machine like no other
- Constantly compacting
- A model (like other GPT ones) that hides thinking traces and thinking summaries, which infuriates me
I've been in the Claude camp for a while, but the way it writes has left me with a a brick for a brain and wanted to see if Astra was as good as they say. Well, I can't know, because in the time it takes for it to actually build anything useful, I've moved to other ideas.
Unbearably, annoyingly slow. I keep thinking I must be doing something wrong.
However, it is not a 'token eating machine'. In fact it uses a third of the output tokens of Opus 5.5, Fable 5.1, or Opus 5.
17k for Astra xhigh vs 61-66k.
The rest still stands, though.
But if I've learned anything is that in a 2 months I might have completely turned around, who knows
https://developers.openai.com/api/docs/pricing?latest-pricin...
One potential deciding point is that Claude still has a $200/mo 20x plan, where, since Sept 11, OpenAI does not and has no ETA for the return.
I downgraded my OpenAI plan 2 months ago to the $100/mo, but my usage has gone way up, but now I can no longer upgrade to the $200/mo plan ("This option is temporarily unavailable"). Thankfully I have 2 usage resets available, but I'll probably be switching back to Claude; I was super happy with Astra but I'm burning through tokens and have 4 days before my next reset.
Its a great release, I will use both heavily.
>> $4 → $2
>> $20 → $10
Do you mean 100% more expensive? GPT 6 is 100% more expensive than 5.6 per your post.
If these price changes mean that coding plans have effectively more usage then that's great, but Codex is surviving on resets from my own experience using it. I was glad to go back to Claude.
This is great, but practically, I'm not going to start working on more side projects.
Perhaps in another 6-12 months I'll be fine to drop down to $20/m instead of $200.
A lot of what I'm doing has pretty expensive build/testing processes between iterations - even on a 40 core machine - so I'm not burning tokens 24/7 like some people may.
I'd guess I'm probably spending >50% of the time running tests & build processes & tooling and the remainder is purely burning tokens.
I also have some internal tooling (that I will hopefully open source soon) that makes LLMs substantially more correct (thus more efficient) - so there's that, too.
There are a ton of use cases that open up with cheaper models.
E.g. extensive security scanning on every PR, quality scans, adversarial reviews etc
Should be B vs A correct?
Else it's confusing
We don't know how much they are bleeding financially, it might just be a front
Highly subjective take
What kind of work do you do, out of curiosity
If they’re subsidizing my usage, that’s great.
In the event of a crash, the investors who put countless billions into this will be still be seeking to maximize their return. Even if it is just pennies on the dollar. Assets (including compute hardware) will be sold, just as they are also sold when any other business fails.
Or maybe a crash doesn't happen. Maybe prices rise to the moon instead and there's nothing we can do to lower them.
Or maybe (just maybe!) a crash never happens and there's never a huge price increase. Prices stay low-ish.
All of these possible outcomes suggest to me that the maximally-sane option that a user can select, today, is to burn it while it lasts. And then, if/when a crash or a massive price increase occurs, just adjust accordingly. (The rest of us will all be in that same boat, too.)
Is there a world where OpenAI starts charging $2,000/month for what we previously were paying $20 for? What are we going to do? AWS could totally jack up the prices for EC2 instances as well, but we've come to rely on that as well.
When they cut prices on luna the first time around they took (literally) millions of users from anthropic.
The "paradox" is when an increase in efficiency which would decrease the use of a resource all else equal, instead indirectly causes more use.
Surely, a large part of the increase of the demand in LLMs is in their intelligence, but to hit the demand models needed to be made more efficient, and labs found that more efficient models, still demanded more usage.
Cache read/write decrease by 50% or similar? That's where most (95%+) of the cost is for agentic coding workloads.
https://developers.openai.com/api/docs/pricing
Performance increases both with larger model (Luna vs Sol)
And with more reasoning (low vs xhigh)
GPT-5.6-Sol, GPT-5.6-Terra, and GPT-5.6-Luna were released in July of 2026.
The first release from the GPT-6 series was GPT-6-Astra. GPT-6-Astra happened on around September 3, 2026, and the previously-mentioned GPT-5.6-* widgets remained available.
Today, September 22, 2026, we now also have GPT-6-Sol and GPT-6-Luna added into the mix.
As I write this, all of the model identifiers I've mentioned are available to select for use within Codex.
I actually preferred 5.6-Terra not because it is technically superior (it isn't) but because it had better instincts to NOT do this stuff.
PS - Speaking of better instincts, have they closed the UI-design gap at all? I keep a Claude subscription just because /design produces significantly higher quality UI design/UI feedback/UI refinement than anything I've seen from OpenAI.
I force OpenAI models to use image generation for design, then an iteration loop until it matches the image gen.
This is frustratingly manual and takes many more repetitions compared to Claude (and especially Claude Design) which "just work", but it's a big step change over the default.
IMO 5.6 Sol had this weird dead zone between medium and high where medium under engineered and took short cuts and high over engineered and ignored instructions it didn't agree under the guise of trying being helpful. The whole 5.6 line was the first release from OpenAI where it felt like reasoning level really mattered and was incredibly finicky.
I haven't felt similar issues with GPT 6 though and am very happy with Astra low/med/high as my default choices depending on the task.
In general, I felt like with 5.6 the effort level did less than previous to make the models smarter and more just increased the complexity of the response. I have a half joke theory based only on vibes that OpenAI splitting 5.6 into Sol/Terra/Luna is where the intelligence split happened and so the effort levels were just like "think harder about the decision you already made". So like if the model decided the earth was flat on low effort it'd just say something like "the earth is flat because the horizon is flat". If it was on xhigh reasoning it'd give you a massively complex answer about how the sun reflects light because of the ozone layer and why people flying in planes can see a curve. In both cases though, adding more effort wouldn't get it to realize the earth was round. It just made the answer about it being flat more complex.
To be clear, that theory is not meant to be taken too seriously. It's not based on anything other than vibes. It's just my way of explaining to myself something I'm frustrated about to myself.
The GPT / Codex models have always been "overengineer" personalities. I prefer that to "I left a pile of race conditions lying around and big gaps in testing" though, which is what I was getting from Opus at times.
But yes both Astra and Sol veer on the side of paranoid. And honestly that's better for team work. For solo work where you just want to yeet something, it can be tiring.
You learn to tame the GPT "personality" on this front by combing over once a week and asking it to find and exterminate pointless tests, clean abstractions etc.
Luna 6 High: https://threejseval.com/models/gpt-6-luna-high
GPT-6 Sol coming soon too.
Also go vote on https://threejseval.com so you can help evaluate how it performs compared to other models!
When 5.6 dropped I had no weekly limits and I could just drive my work with Sol xhigh and things were great. Once limits were back (and maybe token prices changed iirc) Sol was no longer usable (on Pro or business) unless I was ok with 4 prompts every 5 hours, so I had to switch to Terra medium/high. I've used Luna for some really dumb tasks like moving files, renaming variables and whatever other old-school refactors I've needed.
Then Astra dropped and it just uses so many tokens I've only prompted with it once. Now with GTP-6 Sol/Luna I'm not sure what's being said here but most importantly I'm wondering whether Luna 6 is a good replacement for Terra.
Has any other Terra user tried and knows more or less than answer to this?
In 1840 ~70% of the population was in agriculture. That is now ~2%. Things change.
All I want to know is how old is the model and how much does it cost. I can figure out which one I want to use based on that, assuming that newer models are always better.
Trying to convince us there is a difference between GPT-6-Sol and GPT-5.6-Terra or whatnot is ludicrous to the point of being insulting, especially when new models come out every week.
What's unclear to me is who OpenAI thinks they're marketing to with this form of branding. These different models don't really mean all that much to the vast majority of people using their products who aren't developers, and developers aren't helped at all by the way they've been naming said models. Are they merely scared that they'll become irrelevant because Anthropic decided to give their models quirky names like "Opus" and "Fable"?
If OpenAI really wants to give their models names, they should name the generation of model and then have the different sub-models named by purpose or capability level. After all, I wouldn't use Mini for a job that Nano could easily do, and I wouldn't use Nano for a job that the full version of GPT-* necessitates. Similarly, I've had to discover exactly how Luna, Terra, and Sol are appropriate for different complexities and task types. OpenAI could help me skip a lot of those steps and just tell me what each model distillation is good for without causing me to look through their pricing page and make educated guesses. After all, shouldn't they not want me to pay attention to how much they're charging me?
All of this makes the days of frontend framework churn seem quaint and actually preferable.
Edit: Yes, it applies also to subscriptions, source https://x.com/thsottiaux/status/2102463847714247142
If you have something that needs to be done right, might be a bit complicated, up the model size.
You can see this in the pelicans. Big model pelicans are pretty accurate by default. Up the reasoning and only more so, but with more detail. For Astra, it is 105 lines for low, 250 lines for max reasoning.
Small model pelicans will lack the fidelity of a large model. Bits will be out of place etc. For luna, it's 90 lines for low, 150 lines for xhigh.
Additionally the amount of time taken is increased for the larger models. Luna takes 11 seconds on low, and 1:33 for xhigh. Astra is 33 seconds on low, 4 minutes on max.
And naturally, there is the cost. There's some overlap in functionality between luna xhigh and Astra low in the sense that luna really can do quite a suitable job for some tasks. But there are just some tasks that just don't make sense for Luna, even at high reasoning.
The other thing to remember is that sometimes high fidelity isn't ideal. It can lead to overdesigning. My recommendation is to commit early, commit often, and review everything you do, which we've all been doing since before LLMs right?
https://developers.openai.com/api/docs/guides/reasoning?api-...
But effort is basically how much extra internal scratchpad to use and how much extra questions to ask and answer before producing a result, Exploring more hypotheses, validating consistencies, calling more tools.
If you’re happy with your token spend on Astra then keep doing what you’re doing. but if you feel the need to conserve tokens, then you can do that by switching to smaller models like Luna when the task is straight forward.
EDIT: this doesn't say anything about availability on either Azure or AWS. I'm assuming it will show up later, but it would be interesting if it didn't.
Now GPT 6 Luna is even cheaper, and more intelligent, there is no going back... to SOL 5.6 for intelligent layer.
I was hoping for a serious Luna upgrade. It was already cheap enough. This feels more like a price reduction than an upgrade.
That said, if the new Luna is able to handle ultra mode and subagents v2 in codex cli, then at least that’s a win.
I forgot which model degraded in quality as time went by, but let's try out Luna 6 for a few more days to confirm for upgradability.
Design: https://image.non.io/78795662-8bfc-4e14-8d72-3738392aa6b3.we...
All 3 were given the same prompt to dynamically light these and to create the designs as a SPA with page transitions.
Astra: https://html.non.io/annui-astra
Sol: https://html.non.io/annui-sol
Luna: https://html.non.io/annui-luna
Luna gets the button wrong, and in the same way Grok/MiMo did. Looking into it more, it's because Luna actually searched my computer for similar builds, found the ones that I did for grok/mimo, and referenced their files. Astra is still the best by a significant margin in my eyes. Far more polish, better page transitions, effects that aren't overcooked and take into account the page. Better contrast.
> In the Coding Agent Index, Sol improves but Luna regresses: In OpenAI's Codex harness, GPT-6 Sol (max) scores 57 in the Artificial Analysis Coding Agent Index, up 2 points from GPT-5.6 Sol (max), with gains in Terminal-Bench 4.0 (43% vs 37%) and SWE-Atlas-QnA (58% vs 54%). At $2.99 per task it costs ~50% less than GPT-5.6 Sol (max) and sits on the Pareto frontier of Coding Agent Index vs Cost per Task. GPT-6 Luna (max) scores 41, down 2 points from GPT-5.6 Luna (max), with lower scores in SWE-Atlas-QnA (44% vs 49%) and DeepSWE v1.1 (64% vs 66%), at ~60% lower cost per task.
https://x.com/ArtificialAnlys/status/2102462962758033624
Given that they had to discontinue sales of the 20x Pro plan after the Astra release due to compute constraints, I wonder if 6 Sol & Luna are smaller vs their 5.6 counterparts?
At least it sounds good on paper, the the graphed results do give me pause as it seems the lower cost might come from a slightly nerfed base model combined with more thinking, going by the more erratic scoring curves and the lower no thinking baseline score. I’ll have to try it out but I really hope they haven’t nerfed Luna/Sol to make this price point possible!
Does anyone know how exactly these price differences for example between sol6 and sol5.6 translate to codex percentages? In theory it seems like for "high" on both it should result in ~3x more usage. If that is actually the case it would be huge! But all we see is % left and % changes while using and we really have no idea when or how those numbers are being calculated or when they change. So there is a 50% price reduction on API but who knows how the hell that translates to whatever price calculation is used on codex.
Overall, I expect for most people think the winner of today was Anthropic. I personally am preferring Opus 5.5 at medium over GPT-6 Sol Max, in very very early tests. Similar price range, more capability.
But competiton is great, these are solid releases by OpenAI today.
OpenAI/Anthropic meanwhile feel a bit like they're hoping to sell iPhones in a market about to be flooded by $20 flip phones, with almost no channel of their own to do it. And for whatever mad reason OpenAI are now signalling they will attempt to compete on price with flip phones despite their cost of labour, energy, and just about everything else being far higher
I don't see your metaphor to iphones and flip phones. This new Luna model is cheaper than deepseek 4.1 flash, except for cache reads. OpenAI having to compete with China is a much larger economic-political issue that is far larger than just our AI labs.
Then they do a new model launch, issue quota resets all around, and it's a party for 2-3 weeks before things return to normal.
* Prompt caching dashboard: https://platform.openai.com/usage?usage_section=prompt-cachi...
* Adjust reasoning effort and tool availability without breaking cache
For example, I had Fable review Astra’s output yesterday, and it found some issues and fixed them. Passing the fixes back, Astra then uncovered additional issues with Fable’s fixes (and yes, this will go on ad infinitum if you let it, but these were “real” issues).
It seems the big story here is the reduced Luna pricing. It’s a fantastic model that can handle most automation needs (though I still use the big models for day-to-day development).
Nvidia's top AI chip Rubin sells in 72-GPU racks for about $3.5–7.8M. A rack running Xiaomi's MiMo V2.6 Pro generates roughly 150–300B tokens a day, worth about $130–260k at Xiaomi's API price. That's a payback of the infrastructure in a few weeks in theory. After a few weeks or a month, the only cost is electricity, and whatever they make after that is pure profit.
OpenAI and Anthropic are practically scamming people with the token prices.
Astra is the best. Luna is cheapest then it seems like Sol is the middle child like Terra.
Wierd!!
Has anyone else noticed this?
I guess the chinese competition spooked them.
How do you decide what to pick? I mean, I do Platform work on a large monorepo with many different interconnected services, and so I always want the implementation to be "correct".
> GPT-5.6-Sol is retiring. This conversation will automatically switch to GPT-6-Sol
I don't recall OAI retiring a model so early lol. Similar arch?
OpenAI seems really competitive in most areas, and extremely competitive on cost, but still behind on coding.
I kind of hated Astra for it's poor instruction following and stopping all the time plus bad code quality. It somehow feels a bit like some of the popular open models but with a lot more knowledge or peek capability. But it doesn't reach peek that often
Hardly enough time had passed to develop the data to come to a conclusion. Users can take time to build interest.
Funny thing is they very recently also set a real limit per-user/month, so why even limit the models because theyre "too expensive".
Terra ended up just being an awkward middle ground that was not particularly suited for any workload.
Maybe for one-shotting large things Sol is better, but for prod code where I decompose into smaller tasks and read all the code I favored Terra.
Sol 5.6 was still king for architecture/research in my workflow, though.
Hardly enough time had passed to develop the data to come to a conclusion. Users can take time to develop an interest.
Have the frontier labs stopped trying to increase context window size?
Incredible.
Does anyone care about code quality anymore?
This will make me more valuable in the future when everyone has lost the ability to do anything on their own.
I respect that you want to learn how things are done, that is a great trait. But once you learn how its done, you should use the tools to free up cognitive load for more difficult tasks.
Not exactly a low-friction user experience!
Like are you supposed to just somehow intuit, “ah yeah, this task is definitely a GPT-6 Sol Medium task,” or something?
Is this just second nature for OpenAI employees? How are end users supposed to know how to optimally choose a model for a given task? Am I missing something completely here?
Maybe OpenAI can offer an "auto" mode for Codex on the subscriptions, while leaving the possibility of users manually overriding whatever model the router chooses. To me that would be the best of both worlds. The problem is building a competent model router.
Which... fine, I'll take that.
From around GPT 4 results got "Good enough"...I generally try to explain what problem I'm trying to solve, set limitations and boundaries, tell it to ask me questions, have it write up a plan with steps then we take one step at a time.
These new models are starting to feel like iPhone releases where the improvements / feature set feels incremental.
Same on the Claude side which I use for work
I wouldn't be curious to sign up to codex whatsoever these days
These token reset shenanigans are insane
Asking cuz I don't think I'm a bot [pats self], I legitimately prefer the GPT models to Anthropic's, don't like Anthropic's customer service/reliability story at all, and I welcome a massive price reduction. Seems like something I should be happy to get.
If you'd told me I'd be typing this a year ago I'd be skeptical though.
But the reason people say "Claude can't compete" is because Claude Opus has been going downhill since 4.7, and many have found Opus 5 intolerable. Fable is much better, but also much more expensive than OpenAI's offerings.
OpenAI is promising "the Sun, the Moon, and the Stars". The spirit of P.T. Barnum is doubtless looking on with jaw dropped at what is beyond doubt one of the greatest demonstrations of chutzpah, by some of the greatest hucksters, in the history of the human race.
As for any cost based argument, it is immediately invalid because the cost is something that OpenAI fully controls and manipulates.
My previous comment--which suggested astroturfing--was the highest upvoted comment here until it got flagged. Which implies to me that atleast the other remaining humans on this forum see it as well.
26 minutes, 89 comments, upvoted instantly to the top, posted within one hour of the Opus 5.5 announcement. You tell me.
On the other hand, you have previously written: I'll gladly admit I think what these companies are doing is unethical, and I'm sure that biases my thinking toward skepticism. [1]
You have now posted accusations/assumptions of astroturfing and manipulation at least 15 times, without ever providing any evidence. This is in breach of the guidelines, because comments like this poison discussions far more than the comments they're complaining about.
We – of course – want all comments and posts on HN to be authentic. HN is only a place where anyone wants to participate because since the beginning, we've had software mechanisms and moderation practices that detect and weed out inauthentic commenting and voting. We're identifying and dealing with it every day, continually improving the software to detect and remove it. Most of that happens quietly and efficiently in the background without anyone having to see it. When users see evidence of manipulation and report it to us via email, we happily and thoroughly investigate it.
Most of the time, what we find is simply that people are authentically excited and passionate about the topic, which is what is happening here. I understand it can be hard to accept that if you're skeptical about the topic.
It's fine to be skeptical about the topic and you're welcome to express your skeptical views on the topic. People do that every day on HN, about AI-related topics and countless others. Healthy debate is what we're here for.
But you can't keep poisoning HN, by (1) continually posting these unfounded claims, then (2) when users and moderators simply uphold the guidelines, staging a protest by demanding your account be deleted. This is not what people do when they care about a forum's health.
[1] https://news.ycombinator.com/item?id=48220908
> You have now posted accusations/assumptions of astroturfing and manipulation at least 15 times, without ever providing any evidence.
I tried to point out the upvote speed and age-to-comment ratio for this thread look anomalous to me, and that it being posted within the Opus 5.5 release hour was further reason for skepticism. Circumstantial, sure, but I see very little ways to gather hard evidence of astroturfing without being a mod.
> when users and moderators simply uphold the guidelines, staging a protest by demanding your account be deleted.
You're right this was a little dramatic. I think it is just annoying though when two times that I have posted about astroturfing, it has been the most upvoted comment only to get flagged. I guess this is a self-fulfilling prophecy though, as you are right that other human commenters are abound and tend to flag people complaining about astroturfing.
Anyway thanks for the reply. If your read of my account is that I'm more-often-than-not a bad actor, then I will stop commenting here. Seems like it is for the best. :)
> Hopefully it is clear from my other comments that I do try to provide value too.
I agree that you provide value in plenty of your comments, which is why we don't just want to ban/lose you.
> I have posted about astroturfing, it has been the most upvoted comment only to get flagged
People love a conspiracy theory, and on a site like HN that has many people looking at it at once, it's easy to get a large number of upvotes in a short amount of time if enough people find it exciting. We often see off-topic, titillating one-liners or ragebaity comments at the top of threads, and we always have to downweight them to keep the discussion on-topic and healthy.
> Anyway thanks for the reply. If your read of my account is that I'm more-often-than-not a bad actor, then I will stop commenting here. Seems like it is for the best. :)
It seems like you're well intentioned. You have your concerns about A.I., as many do and that's fine. You're still very welcome here. Just please try to believe that many or most of the people who are enthusiastic about A.I. are as sincere in their positivity as you are in your concern.
fyi, i flagged it because it is boring reading and against the rules.
if you suspect astroturfing, flag the comments and contact the mods.
(complaining that your complaint got flagged is also tiresome. contact the mods. "@dang" doesnt work, use the email.)
I respect you for replying here though, and yes I get that HN forum standards would suggest flagging my previous comment. But it is just sad to see a place used to be so vibrant get manipulated because of how much weight it holds for us in the industry.
And yea sure, I could go and flag all the bots and message Dang. But probably time to stop shouting into the void. :)
But do you really think people were so excited about cheaper versions of Astra that they were just waiting around to comment the instant this was posted? More than two comments per minute? All the initial comments were really similar too: brief one liners celebrating the cheap prices.
I think AI right now is a sort of Rorschach test. What it is clearly revealing to me is that I don't trust organizations with enormous financial incentives to not manipulate public opinion. So I see bots everywhere. :)
> what the fuck
yes, i think so.
because, unfortunately, complaining about bots (or astroturfing, or whatever) doesn't stop them. so we end up with threads that have both the potential bot/astroturfing/whatever activity and complaints, which further drowns out any interesting comments.
Anyway these comments were made when this thread was in an earlier state. I agree that it has gone on to be more "organic" looking. That doesn't exclude it initially being manipulated to the top, in my mind, but certainly they aren't carpet-bombing with only booster comments.
Also, I don’t see that much astroturfing here? (And I tend to see it a lot on HN.)
I would agree now that the thread has recovered to a more interesting state, but how it first looked--combined with it being posted right after Opus 5.5 announcement--look questionable to me.
I see long massive pro apple threads. I don't get it at all. As in; literally don't understand what Apple is good for. But I have friends, family irl who love apple so I know the sentiment exists. I therefore accept that many HN users are similar.
Many people really truly are happy to see another model drop and are excited about progress etc etc. Surely you've met such ppl in real life. Well, they're here too (I'm one of them fwiw)
I wouldn't argue there aren't real humans excited for this drop. It was just all the circumstances around it--the comment speed, the upvote speed, the initial uniformity of what people were saying.
Anyway, thank you for the moral reminder.
This is why HN has been on the down hill in quality and those that care to highlight that are being punished, while the astro-turfing, gaslighting and Show HN self-promotion slop continues.
Otherwise, yes, we agree. Although, given the other replies to this, there are clearly those who disagree who appear to be smart and level-headed.
Anyhow I've learned my lesson now.