This squares with my, much much, smaller OpenRouter usage. It’s just incredibly unreliable and you are forced to pin providers and even then it can be a crapshoot as the author found.
OpenRouter sells the idea of swapping being commodity providers but it couldn’t be further from the truth. Provider A is often not swappable for B or C (again, as this author found). It can be crazy-making as you sit there thinking “OpenRouter has no clothes right?! Am I the one that’s wrong?”.
I love the _idea_ of OpenRouter and maybe Stripe can improve this situation but the only sane way I’ve found to use it is to tightly pin providers to the point I wonder if I should just use the providers directly.
Without pinning you are in for a world of hurt and unreliability (varying model capabilities, speed, etc).
> just incredibly unreliable and you are forced to pin providers and even then it can be a crapshoot as the author found.
But that's the intention right? Even the name implies they just send stuff around for you, and if you want to control the routing, you'd lock down providers. I don't see how they could build what they wanted to build, and not have it end up unreliable if you freely round-robin between providers, it's bound to work exactly like this.
> I love the _idea_ of OpenRouter and maybe Stripe can improve this situation but the only sane way I’ve found to use it is to tightly pin providers to the point I wonder if I should just use the providers directly.
This is quite literally the point of OpenRouter. A unified interface, so you can easily switch providers without changing a ton of code which using providers directly would most likely mean, as there are slight differences between them. And the providers all run different weights, so of course quality/performance will differ among them.
I guess OpenRouter is a bit like Amazon, in that they're just routing stuff around for you, but to actually find the good and usable stuff, you need to focus in on what providers/manufacturers you know are good, and stick with those. Still, the unified interface helps you to shop around and try different ones when you want to.
1. Yes, but if i can't rely on openrouter to route to providers to give me the best price and the best uptime across a number of providers, and it instead gives me inconsistent results, then I will not use them.
2. As long as the apis use the openai spec, it should be fine? And what makes you say the providers are using different weights? The blogpost says the exact opposite?
3. Great example, if amazon does not lead me to good products, I will stop using it, and this is why I dislike amazon. There are so many crap products, and the search seems to try to push crap products instead of what i'm actually looking for at a good price.
You take a bunch of providers with not great uptime, put them in a pool and now you get great uptime - but it doesn't work if it's at the cost of shitty performance or failing toolcalls.
> 1. Yes, but if i can't rely on openrouter to route to providers to give me the best price and the best uptime across a number of providers, and it instead gives me inconsistent results, then I will not use them.
OpenRouter does reliably route to your specified model and provider, otherwise it'd pretty much be fully broken. Parent is complaining about the auto-provider chosing, not that all providers are unreliable.
> 2. As long as the apis use the openai spec, it should be fine? And what makes you say the providers are using different weights? The blogpost says the exact opposite?
In theory, yes. In practice, no, there are differences. Ollama, llama.cpp, vLLM and SGLang all say "ChatCompletionRequest" compatible, but the devil is in the details, they don't have 100% the same request/response schema across all compatible models.
> 3. Great example, if amazon does not lead me to good products, I will stop using it, and this is why I dislike amazon. There are so many crap products, and the search seems to try to push crap products instead of what i'm actually looking for at a good price.
Yup, makes sense! If you're unable to find models when you use OpenRouter, it makes zero sense to continue to use OpenRouter.
> You take a bunch of providers with not great uptime, put them in a pool and now you get great uptime
Huh? That's not how it works or does it make sense, nor have I've seen anyone use OpenRouter like that.
if you want to control the routing, you'd lock down providers
I don't want to control routing. I want the model to work how the giant, prominent "BENCHMARKS" section says it works, not randomly have a 100x error rate.
If OpenRouter is a marketplace to pick a provider while avoiding huge problems, it is terrible at that job. It surfaces literally none of that info in the top-level list, the graphs below are mislabeled and useless at best, and doesn't notify you of this horrifying situation anywhere, even in passing. There's not even a way to compare providers, AFAICT -- you can only compare models.
This is quite literally the point of OpenRouter.
Their tagline is "better prices, better uptime, no subscriptions". The first two of these directly and inherently contradict your understanding -- neither would be possible if OpenRouter was just a fancy way to change something in their GUI rather than changing the target url of your gateway.
> I don't want to control routing. I want the model to work how the giant, prominent "BENCHMARKS" section says it works, not randomly have a 100x error rate.
Why do you care about the public benchmarks at all?
The way companies and effective individual developers use OpenRouter, is that you first create your evaluation framework/benchmark, for your specific tasks and use cases, and make that real easy to run and use various models and providers with it.
Then you run this to gather data. Then you use said data to figure out what works and what the quality/cost tradeoff you want to make is. Then you lock that down in production while you keep iterating on your benchmark to make it match with real-world use cases and keep adding the new models that pop up.
I don't think anyone serious is just willy-nilly making individual requests against OpenRouter and similar platforms, get a "feel for a provider" then use only that provider. Not only would it be wildly inefficient, but also you need hard numbers to compare so you can make informed choices.
For this process and workflow, OpenRouter is great, because adding/changing providers and models is essentially changing two strings, rather than having a adapter for each platform you want to try out.
If you just want best accuracy requests from SOTA models for your agent you run locally or whatever, then don't use OpenRouter, it doesn't make much sense, but use the provider the model maker has available, as almost all of them run their own endpoints.
If openrouter is only for people who "make their own benchmark" in a mission to roll something that's usually made by scientists with large budgets, it should say so and thus fade into deserved obscurity.
I don't think anyone serious is just willy-nilly making individual requests against OpenRouter
Despite your confidence, that is indeed the basis of this massive corporations entire business plan.
If you just want best accuracy requests from SOTA models for your agent you run locally or whatever, then don't use OpenRoute
If OpenRouter is only for bad accuracy, they should say as much and fade into deserved obscurity.
Ouch, you really hurt my feelings there, good job. Speaking of which, you never rotated your account? Imagine being so self-centered you think karma points on the internet matters one iota...
You'd be surprised. Not enough teams still have their own evals. India[1] and the US[2] is my experience. To get many of them to understand the benefit of putting a couple of people to do data labelling and write a couple of verifiers for just a few days every few months was so difficult. Many folks have understood it all wrong and made allotments like "big model for this task" "small model for this task". Small was sometimes parameters, sometimes brand version number, or sometimes because it has "mini" in its name.
Proper LLM adoption beyond fucking around with claude code or github copilot is so low and is only going to go up as people figure it out. I also think the new cloud agents thing might accelerate adoption among these companies, since it's a bit more plug and play. But they are more likely to be stingy about it, so I'm not sure about the high margin expectations of certain model families.
[1] not witch, but BFSI. Surprisingly parts of witch companies have it down to science already, they embedded openai or anthropic or startups like devrev more than a year back.
I treat OpenRouter as a central point to access specific providers. Pinning is a given for that use case.
I’ve tried using un-pinned models and the experience is exactly as you described: Some providers are so unreliable that the majority of requests fail. Some providers do weird things like abruptly end the response (which I get billed for and have to re-submit). Some providers are clearly running heavily quantized versions of the model because their eval performance is terrible. Some providers advertise features on OpenRouter but will reject those requests when submitted to their API.
This (by definition) the reality of every normalization/proxy layer, standard, etc. There’s always a cost for flexibility/freedom. What’s traded is when and by whom it’s paid.
I think Openrouter is great for quick testing, getting a "feel" for the model, or very quick integration jobs where you want to try out as many models as possible.
For truly production use cases, use Novita, Fireworks, Toghether or something of the sort.
> to the point I wonder if I should just use the providers directly.
How many account credentials, balances, and tokens do you want to maintain? Even without automatic failover services such as openrouter are still incredibly useful.
Personally I pin a single vetted provider in the interest of minimizing risk.
You know you can define your own provider filters and orderings, right? The filters and such are not _that_ advanced, but it might do what you need if you haven't already tried that.
I only use OpenRouter to benchmark models. For my production app, I have direct integrations with OpenAI, Anthropic, Google, xAI. Even just input caching is enough of a reason to do that, assuming you are trying to build the fastest and most cost-efficient product possible. I don't understand how people supposedly run production apps using OpenRouter unless those two variables somehow don't matter to them.
Yes to all this but more. The thing that made me leave and go to a single provider was token caching. I have to keep blocking providers that don't properly cache. I see performance tank and then I look in the logs and a new provider has been rotated in and every call to them is uncached because they are clearly broken. This has happened a few times now and essentially destroys cost savings (these providers also often have terrible quality). Don't they monitor for simple things like this? Their own logs show how clear this pattern is for some providers. Simple cache % stats would allow them to block providers nearly instantly.
The best part about OpenRouter is 200 OK is probably hardcoded into their responses.
I used to get content: "" all the time and I used to triple check my code to see if I was doing something wrong until I realized most AI providers in general have vibe coded their infrastructure as well and it is just a futile attempt to even fight it.
Providers probably serve quantized versions without disclosing it. Which is a real shame, because for certain tasks I would be perfectly willing to trade accuracy for cost. But, unfortunately, it is impossible to explicitly choose how quantized do you want your model to be, unless you are running it yourself on your own (or rented) hardware.
BTW, does anyone knows if LLM Gateway suffers from the same issues? Currently looking at trying it, but haven't got to it yet.
> Providers probably serve quantized versions without disclosing it
It should be OpenRouter's responsibility to protect you against it, by regularly benchmarking providers and giving you the control to avoid bad providers.
In fact, that's a big opportunity for them, since it justifies their place as a middleman between users and inference providers.
100% agree. The key issue is that users think they are getting results from a specific model configuration and they are clearly not, which is fraudulent.
If OpenRouter wants to succeed as a business, they need to be auditing the providers they connect to (i.e. benchmarking) and removing fraudulent ones from their service.
I mean that is addressed as well later, even when they do declare the quantization, doesn't mean you'll get better performance than the lower quantized one. I'm guessing they're doing something similar to what stadia was doing, saying you are playing games in 4k, because you're getting a 4k stream, but the game itself is running in 1080p
I'm happy with OpenRouter. They're good at routing requests & spend tracking/management. You should just expect to directly own your relationship with the downstream model provider.
Source: Enterprise customer doing $XXM annual run rate of inference spend on their platform
This is very useful information and comes at a perfect time! I use Openrouter for my newly released running tracker (I use it for live coaching and post-run debriefs). I've benchmarked a bunch of models over time to evaluate their aptitude for this specific task, and have noticed that sometimes a model can underperform for seemingly no reason. I'll be sure to include model providers in my benchmarking suite going forward!
Open router have one more annoying disadvantage: credits you buy expire in around 1 year +/- ~30 days. They do sent warning email 30 days in advance, but then just make them disappear.
If you for some reason had unusee balance and forget about it; its just gone.
Oh, interesting, I feel like a year so ago I read something posted from openrouter team that they write that credits expire, but that they in actuality don’t expire them. But maybe I’m misremembering, or perhaps that’s changed for the worse in last few months :S
I noticed this with DeepSeek Flash and DigitalOcean. They are doing something seriously wrong when serving that model; we need something like an SLA, but for intelligence; it's almost fraudulent behavior.
<Rant>
I have been working on a product for months and for 1 of the specific models have become the number 3 user and very likely soon the number 2.
Its a google model. The edge cases are crazy to deal with and have taken a long time to find. I am also finding that since it has no fallbacks but no published rate limit I am single handedly taking the model down on what I thought were a reasonable amount of request. There is no other fallback that isn't google. I'm worried that with sustained usage from my users on launch in a few days. Clearly google can't be making that much money on it it im like 10% of the usage and the top 2-3 user of it spending several hundred a month in pre launch testing.
So why would they care to help a small time start up? I will have to jump to a cost effective different model and find the footguns all over again at somepoint.
Not sure where im going with this but just wanted to let others know
</rant>
Building a product that depends on a Google product seems like folly at this point. They are notoriously unreliable about keeping products alive. I hope you don’t get burned.
Me too, im a bit nervous because I also have evals proving that flash models are actually getting worse at specific task on newer versions and one of my prime use cases peak around flash 3 and tanks in 3.8. So it would be very very horrible for me if they kill old flash models on open router.
Yay google, love living in fear of a vendor that they will rug pull me at any point because they can.
For a Google model you might be better off directly using their API, which is unfortunately far more painful than using OpenRouter, but it might at least be more reliable for your product.
I think the rate limit issue happens because of OpenRouter sending so many requests to Google.
"OpenRouter runs per-provider benchmarks on the same model: GPQA Diamond and TAU-Bench Airline (a tool-calling task)." why I haven't never seen it? Click on the link brings nothing! Benchmark is only for the model. Per provider is a performance matrix (latency, throughput)
What did I miss?
--
Update: oh, that is the AutoExacto Benchmarks! Now I see it.
The author’s comments on vision providers is especially interesting. We saw that most providers don’t provide native video url support, have high-variability in vision performance (likely due to the fact that they’re serving different quantization levels behind the same model id).
If you’re building vision-native apps, there are so many footguns in vLLM/SGLang serving configurations, let alone the routing/orchestration in providers like OR, that leave the user more confused about the model’s capabilities.
Thanks everyone for the feedback here. Some of this we are aware of, some of it we aren't. Some we can fix, some of it is inherent to inference (and we in fact improve the situation dramatically).
Philosophically, at OpenRouter we are trying to do two different things, that are sometimes at odds with one another:
1. Let you use a lot of capacity across a lot of providers, in a way that "just works" and you don't need to worry about it.
2. Have a huge variety of inference available so you can pick radically different price/performance tradeoffs, data policy decisions, geographic destinations, inventive hardware, etc.
These are inherently odd bedfellows, and we are still very much improving how we can make both of them true at the same time.
Some quick thoughts on the article itself:
1. Benchmarks: YES! Providers benchmark differently. We run benchmarks on the live endpoints continuously, monitor the median performance, and kick providers out of the default routing pool if they vary by more than a standard deviation. We work hard (and continue to invest) to make sure that providers serving sub-par inference can't game the system, and that our routing actively avoids them. So the chart is accurate (it's our chart) and it actively influences our routing decisions!
2. That is bad and we will fix it. Sorry.
3. When we on-board providers we run essentially the same test as the author did to verify that the param is working as expected. If it isn't, we don't launch the provider. However this is not one we are running constantly in production. We are working on making this more robust in general and I do believe is fundamentally solvable in a way where it will "just work".
4. We 100% agree that users should not filter by quantization. It's a bit of a legacy concept in general; there is a huge amount of code between "model weights" and "inference API" and in almost all cases quality degrades in that part of the stack, NOT in the model weights themselves.
5. Hmm...we will dig in here. We monitor tool calls in real time and route around providers that are regularly mis-parsing tool calls. So you should get a very low rate of these in general. Another area we have invested a lot in: https://openrouter.ai/docs/guides/routing/auto-exacto
6. We will dig in here as well. I'm surprised this is happening frequently enough to be noticeable. We eat the cost when the finish reason is an error, but not when it is "stop". Perhaps we can expand our "insurance" program: https://openrouter.ai/docs/guides/features/zero-completion-i...
7. Will investigate.
8. We attempt to heal these, but obviously missed some. Will fix.
9. We do not rate limit by IP. Would love some more information here, as that is very surprising.
10. Ugh. That sucks. I'm sorry. We are introducing QoS tiers for production apps, which will address a lot of this.
toven from openrouter, leading the team working with our providers - very interested in some of the things found in the report, I dug in to the image failures specifically, and in that case we have data showing deepinfra was correctly parsing images when the endpoint went live in july, but today fails those tests. we'll work on testing images and reasoning effort etc running constantly as chris mentions in point 3.
Hmm, let me check. That certainly seems wrong. Can you send me an email w/ your email so I can look into it? im cc at openrouter.ai. Or DM me on X? x.com/cclark
My ticket number is #107156 (opened 9 days ago). The point is that the Discord is full of people not getting responses, and I'm not even the one waiting the longest. The only help they get is from the AI assistant in the server saying don't bump the ticket or you'll be punished by getting moved to the back of the queue. So there's nothing anybody can do but wait who knows how long.
One thing to note about the first graph: nobody is doing as well as the first part on tool calling, and it's not close.
This might be the fault of the other providers, but it's probably just something slightly different that the first party does with the model inference program than anybody else, and that's not sure to weights it's due to vLLM twiddling (or whatever) and probably becuase the first part actually uses their own customized inference program rather than the standard methods that all the third party providers use. This isn't nefarious, it's just the challenge of these sorts of stochastic systems.
Having been in science for decades now, and seen benchmarking across many different fields, these results are completely expected for me. LLM serving is not mechanical, it's hard to get right and has lots of unknown footguns. Even something as extreme as scrambling a matrix will still likely get results that are nearly as good as normal, and if there's a bug deep in vLLM or the tensors metadata that results in that, then it's going to be pretty hard to find unless you're an active researcher with knowledge of the particular model you're running inference on. I kind of doubt that's happening here, but maybe!
In the scientific literature, when benchmarking methods, everybody's own method performs best in their own hands. Some attribute it to researchers gaming benchmarking for publication purposes, but I think it's just what we see here: the people who made a method are just the best at using it because they know all the quirks and use it best.
Programmers are not used to thinking with that nuance, and jump to conclusions about lying about quantizations, etc., but this is really just an unavoidable part of AI/ML methods: when things aren't perfect they're still pretty good and it's going to take the model creator to truly debug it. At least until the open weights ecosystem gets a lot better at ensuring reproducibility, and model cards are nowhere detailed enough for that to happen yet.
I also believe some providers fallback to another model entirely. I was recently using Kimi K3 and saw that some requests had no reasoning trace whatsoever. Unsurprisingly, those requests were routed to the less reputable providers (Sail Research).
I've had opencode go + opencode work reliably, though I'm skeptical of how robust their data security claims are in practice because they suddenly blocked accessing Deepseek unless you were okay with the data going to China where true data privacy for something like that is illegal, which makes me wonder where it went before, which weakens my trust. It's also a lot less useful now that Deepseek is so much more expensive.
I was looking for an LLM gateway and saw that the most popular one had just had a massive supply chain attack, so I wrote my own. Took about 2 weeks and initially I wrote it as a provider for pi coding agent. I connect to moonshot, qwen, Gemini, zhipu, anthropic, deepseek and OpenAI. I use models.dev to load model and pricing info. Adding new providers is pretty easy because I have a standard internal format and each provider has an adapter that translates between my standard format and that required by the provider.
Fireworks has high variance amongst models and while some are served correctly; many of them are junk / broken and degraded and it seems like they don’t even know; because even running 1k MMLU Pro questions would flag it very quickly.
I've been using Merge AI Gateway and it's been useful so far. They tend to add new models quickly, and support has been responsive. https://gateway.merge.dev/
Openrouter is useful for quickly testing various models with just one API. In development, it's useful. I would not run it in production tho' for all the caveat mentioned. Go to the first party provider directly, it's cheaper usually. And the cost to rewrite to use their API is usually noting (you can even have both and a feature flag), especially if you just vibe code it.
I've used Requesty (https://www.requesty.ai), let's you pin down providers and build your own routing policy so you at least somewhat know what to expect.
There are none, this isn't a problem specific to OR as much as it is a problem with serving LLMs in general.
If you use any other meta-provider that routes your requests to third party providers, you'll likely face the same issues. If you try using any of those providers directly, you'll likely face some of the same issues as well, except you won't have the option of quickly swapping to a different one and taking your credits with you.
Extreme variance in quality and feature support per provider is probably the biggest obstacle holding back adoption of open weights models.
Just use a single vendor. Literally nothing wrong with that, and you avoid the complexity of both n-1 of the vendors (leaving you with the highest quality vendor) as well as the issues with the aggregating layer.
Not sure why people are drawn to this particular blunder. The promise of vendor neutrality maybe? I'll take working product over vendor-neutral slop anyways.
They used a closed list of 3 vendors in prioritized order, and got 429ed out of two of them, while the third one stopped serving the mode.
This is less of a problem if you're running an agent locally and routing your problem to OpenRouter - you can pin to one or two models for consistency and just switch models when something goes bad. But the article is specifically about production traffic.
Never before have I heard this sentiment, NGL. Vendor-neutrality has been an OS(/FLOSS) darling for, well, the whole time.
RE:"single vendor", if this post is to believed then you might have picked one that has 100x the tool calling errors for the next SoTA model, if your single vendor serves the next SoTA in the first place. It also completely erases the notion of competition driving down prices -- that would only hurt you in the short term, but obviously would ruin the whole ecosystem long term.
How does it erase the notion of competition driving down prices? Endpoints are largely compatible, so the code change required to switch from one to another is trivial. Don't load 6 months' worth of credit in an account, keep it tight. There's fairly little lock-in.
The most significant lock-in to me isn't even something you mentioned, but rather it's model related; I personally put a little time into trying to optimize my prompts every time I change models, as they all have their own unique... flavor.
As for tool calling errors, it seems like first party providers are among the best, I got the feeling that's what he was suggesting, though of course that's why you test. You can also go directly to Together.ai or whoever else you please.
Like others have said I think Openrouter seems neat for testing, but even just as a hobbyist I've been drawn to go direct to particular providers due to irritating little issues that I now see just weren't me.
Until yesterday I used openrouter mostly as a portal to the big providers: google, anthropic, openai, without having to maintain keys and accounts and credits for all of them separately.
Yesterday I wanted to do a bit of benchmarking a prompt across multiple models. Small requests. Outside the big providers, the experience became awful. This explains that experience.
Very informative article, but I'd say all of this is in favor of OpenRouter, not against it. It gives you full flexibility, not just of models, but of providers. That's the offer!
As long as you are constantly diligent in staying on top of what each provider is doing that may impact you.
It reads to me an argument for self hosting, maybe a less capable model to deal with smaller compute resources, but when starting to use that model and inference software to build a benchmark you can easily rerun when you change things to observe the impact of the change.
Maybe that’s a similar level of diligence but they feel different to me.
I've had the same experience in my personal use. Unreliable output, thinking token leaking into the conversation... Time to first token and tk/s also vary wildly per provider.
The worst case of hallucination I had, was DS v4 flash switching to Italian mid-conversation and impersonating a podcast host for no reason.
LLM ops is not trivial. The systems for running inference are very complex, and running across multiple GPUs and nodes adds tons more complexity. And when LLMs are run incorrectly, they still work, just not at optimal performance. Even noticing that something is wrong is not trivial, and finding the problem is far far harder.
So I guess I shouldn't be surprised at all to see these benchmarks, but still I am!
There are such huge economies of scale with batched inference that it's clear this sort of service will continue, but it has a lot of growing up to do. Even AWS Bedrock has a Claude that feels different to me, but I haven't had a chance to do actual benchmarks that would show that.
I do use OpenRouter for my personal use and this matches with my experience. I still use it because of it's top up model providing a way to not get surprised by out of control costs and being able to switch models with one account, but it's very noticeable at times.
The models I've mainly been using recently are GLM 5.3 Flash and GLM 5.3. While obviously all these models have some variability, GLM 5.3 Flash feels like it oscillates between "I can't believe it's not Sonnet", but it costs a fraction of that and "This feels like I'm back using GPT-4, why am I even bothering with an LLM?".
exactly. you can specify a provider in your request to openrouter, or, better yet, use the :exacto endpoint to make open router automatically choose one that is good at tool calls.
> Reasoning models sometimes put everything in the reasoning field and hand back content: null, finish_reason: "stop". 345 completion tokens, HTTP 200, nothing to show the user.
This is actually expected behavior. No content and no tool call is the same as content-only: the agent decided it's done. Anthropic has done this for a while.
This is an issue self hosting as well. There’s a lot of footguns that give you slightly bad results.
I wonder what tricks one could use to ensure the model is actually performing on par with the reference api, like matching seeds or running exact benchmarks.
Under privacy, ignored, I put providers that always suck. Digital Ocean, OpenInference, DeepInfra, AtlasCloud, and Alibaba (that one is different, it doesn’t suck in the same way as the others with a dumb model, but instead it’s heavily censored and doesn’t like being used as RP model in Skyrim).
Yea I can believe all these. I've personally have been having issues on these points:
"200 OK, no answer" - insane that openrouter's main feature is literally a fallback and streaming doesn't support 200 no content to fallback to another provider or smth.
"rate-limit by IP"... now it kinda makes sense why deepseek v4.1 flash rate limits me on prod but never seems to happen on local. Makes you have to basically pin Deepseek as provider, since I've never had 429 error on them
Wow, that explains a lot. I was using OpenClaw with open router, but stopped because how inconsistent the results were over the same prompt. I thought OC is at fault, never would have guessed that there are differences between providers.
The way that OpenRouter frames it with having calls routed to whatever provider is handy encourages you to think that they’re all the same in basic functionality. Really sours me on OpenRouter.
exactly. all of these people complaining are using it wrong. you cant just expect a layer of abstraction to perfectly fix all the layers below it! plus, open router does a great job of handling a lot of the provider problems, but it literally cannot fix the output of the provider
could not agree more one pet peeve of mine is that using strict json output also does not work on all the endpoints of the models so I have to go through them one by one try it out and then only whitelist them.
I couldn't agree more. This articles describes how frustrated has been OpenRouter experience. At the end of the day, I ended up configuring to use the owner provider of each model I needed to use.
I never tested this but always suspected it. When I learned that providers differentiate themselves on how they optimize and host the model (otherwise, why would you choose one over another?), I figured some were less "give me the best possible experience" than others.
I use Fireworks for a production app. Much lower volumes though. My only gripe is their serverless offering for reranking has only 1 model which is Qwen 8b and it's quite expensive.
I believe OpenRouter addresses this with sticky routing and keeps a conversation on the same provider and falls back if that provider becomes unavailable via an explicit session_id to keep requests together. You’re right though that upon a failure cache is not portable and incurs additional costs.
great post, I just put $100 in credits on open router to try different models, I think maybe using the "real" provider is the way forward once I've spent my credits and decided which one I want (for now).
Be sure to spend them! Another great feature of OpenRouter is that they will take your credits after a year. They just delete them. A ToS-legalized theft if you ask me. (Even if you are still actively using those credits.)
As of right now, they only expire credits if you don't use the account at all for a year. This is different from, say, OpenAI, where your credits expire unconditionally. That said, they do deserve the right to expire them regardless, so this could change in the future.
I noticed some strange behavior when I was setting it up to avoid providers that collect data, hence blocking default providers. Definitely it is worth filtering well tested providers, which is a feature OpenRouter provides.
The article that I read addressed this point specifically. "Trusted" providers do not maintain consistent performance, and do not have the same performance across different models.
Another thing to be aware of -- apparently their billing limits don't work. Somehow someone was able to use my key from Singapore. It had a $10 daily limit and they were able to spend $100. Not only that, it got me blocked from all models for the frontier labs.
On top of that, near as I can tell, there are no protections for your API key. No restrictions by country, IP, etc...
Just chiming in that I've never had this experience with them. I set, hit, and depend on those limits regularly. I would be wildly interested in reading what their support analysis had to say about this situation, if such a postmortem was done.
So far it was just "investigating". From the logs, my best guess is they sent all the requests at approximately the same time, and whatever limiting they do doesn't react fast enough.
I get it, the API key was my responsibility, but, setting the dollar limit is exactly the guard they suggest against that.
I think that's only on the enterprise plans? So no real protection for those of us using open router personally? Or did I miss other protections somewhere?
Wait WTF?! I thought performance was, well, performance, not efficacy! OpenRouter's UI in this section is remarkably broken and unclear at the same time, and I have no idea where the author got those clean charts. "AutoExacto" numbers (completely meaningless name) are available for 6 providers, two of which are the same provider, and none of which are even in the top half throughput-wise. When you click "+28 more providers", it just shows a clearly broken modal. And the the next two graphs are even worse. Really all of these graphs are pretty and completely 100% useless.
I feel like it's absolutely insane that some providers serve the same model with much less efficacy. That doesn't make sense to me. What's going on?! I'm suddenly feeling intense shame for having routed all my non-subscription usage through them so far, and honestly some white hot anger that they would blatantly lie about something so important.
We run benchmarks against all of our endpoints, in production. That first chart that the author shows is in fact our live benchmarking data. If providers underperform, we kick them out of the routing pool. That is why we run those benchmarks. Performance
And errrr...yeah...that auto-exacto performance chart is both 100% useless, and totally unclear. We will get that fixed. But under the covers it is doing a lot of valuable work! https://openrouter.ai/docs/guides/routing/auto-exacto
As expected vibecoding bros cannot even read the manual properly.
It is pretty trivial to pin a single provider for a model. Better yet, instead of calling the model directly, use presets instead. You can easily change the setting on openrouter without having to update your app every time.
The article mentioned they pinned the provider and model and the result was bad as well with their own shenanigans. That part was in the end of the article so maybe you missed it.
It's good to quantify the extent to which a lot of the stuff on there is just vibe-hosted. But I think it was always pretty apparent that this was the case. You've got numerous providers all running the gamut of:
OpenRouter sells the idea of swapping being commodity providers but it couldn’t be further from the truth. Provider A is often not swappable for B or C (again, as this author found). It can be crazy-making as you sit there thinking “OpenRouter has no clothes right?! Am I the one that’s wrong?”.
I love the _idea_ of OpenRouter and maybe Stripe can improve this situation but the only sane way I’ve found to use it is to tightly pin providers to the point I wonder if I should just use the providers directly.
Without pinning you are in for a world of hurt and unreliability (varying model capabilities, speed, etc).
But that's the intention right? Even the name implies they just send stuff around for you, and if you want to control the routing, you'd lock down providers. I don't see how they could build what they wanted to build, and not have it end up unreliable if you freely round-robin between providers, it's bound to work exactly like this.
> I love the _idea_ of OpenRouter and maybe Stripe can improve this situation but the only sane way I’ve found to use it is to tightly pin providers to the point I wonder if I should just use the providers directly.
This is quite literally the point of OpenRouter. A unified interface, so you can easily switch providers without changing a ton of code which using providers directly would most likely mean, as there are slight differences between them. And the providers all run different weights, so of course quality/performance will differ among them.
I guess OpenRouter is a bit like Amazon, in that they're just routing stuff around for you, but to actually find the good and usable stuff, you need to focus in on what providers/manufacturers you know are good, and stick with those. Still, the unified interface helps you to shop around and try different ones when you want to.
2. As long as the apis use the openai spec, it should be fine? And what makes you say the providers are using different weights? The blogpost says the exact opposite?
3. Great example, if amazon does not lead me to good products, I will stop using it, and this is why I dislike amazon. There are so many crap products, and the search seems to try to push crap products instead of what i'm actually looking for at a good price.
You take a bunch of providers with not great uptime, put them in a pool and now you get great uptime - but it doesn't work if it's at the cost of shitty performance or failing toolcalls.
OpenRouter does reliably route to your specified model and provider, otherwise it'd pretty much be fully broken. Parent is complaining about the auto-provider chosing, not that all providers are unreliable.
> 2. As long as the apis use the openai spec, it should be fine? And what makes you say the providers are using different weights? The blogpost says the exact opposite?
In theory, yes. In practice, no, there are differences. Ollama, llama.cpp, vLLM and SGLang all say "ChatCompletionRequest" compatible, but the devil is in the details, they don't have 100% the same request/response schema across all compatible models.
> 3. Great example, if amazon does not lead me to good products, I will stop using it, and this is why I dislike amazon. There are so many crap products, and the search seems to try to push crap products instead of what i'm actually looking for at a good price.
Yup, makes sense! If you're unable to find models when you use OpenRouter, it makes zero sense to continue to use OpenRouter.
> You take a bunch of providers with not great uptime, put them in a pool and now you get great uptime
Huh? That's not how it works or does it make sense, nor have I've seen anyone use OpenRouter like that.
If OpenRouter is a marketplace to pick a provider while avoiding huge problems, it is terrible at that job. It surfaces literally none of that info in the top-level list, the graphs below are mislabeled and useless at best, and doesn't notify you of this horrifying situation anywhere, even in passing. There's not even a way to compare providers, AFAICT -- you can only compare models.
Their tagline is "better prices, better uptime, no subscriptions". The first two of these directly and inherently contradict your understanding -- neither would be possible if OpenRouter was just a fancy way to change something in their GUI rather than changing the target url of your gateway.Why do you care about the public benchmarks at all?
The way companies and effective individual developers use OpenRouter, is that you first create your evaluation framework/benchmark, for your specific tasks and use cases, and make that real easy to run and use various models and providers with it.
Then you run this to gather data. Then you use said data to figure out what works and what the quality/cost tradeoff you want to make is. Then you lock that down in production while you keep iterating on your benchmark to make it match with real-world use cases and keep adding the new models that pop up.
I don't think anyone serious is just willy-nilly making individual requests against OpenRouter and similar platforms, get a "feel for a provider" then use only that provider. Not only would it be wildly inefficient, but also you need hard numbers to compare so you can make informed choices.
For this process and workflow, OpenRouter is great, because adding/changing providers and models is essentially changing two strings, rather than having a adapter for each platform you want to try out.
If you just want best accuracy requests from SOTA models for your agent you run locally or whatever, then don't use OpenRouter, it doesn't make much sense, but use the provider the model maker has available, as almost all of them run their own endpoints.
10month old account with 20k karma. Low value rubbish postings as a professional user. Sad.
Proper LLM adoption beyond fucking around with claude code or github copilot is so low and is only going to go up as people figure it out. I also think the new cloud agents thing might accelerate adoption among these companies, since it's a bit more plug and play. But they are more likely to be stingy about it, so I'm not sure about the high margin expectations of certain model families.
[1] not witch, but BFSI. Surprisingly parts of witch companies have it down to science already, they embedded openai or anthropic or startups like devrev more than a year back.
[2] again not high fly SF companies, BFSI.
I’ve tried using un-pinned models and the experience is exactly as you described: Some providers are so unreliable that the majority of requests fail. Some providers do weird things like abruptly end the response (which I get billed for and have to re-submit). Some providers are clearly running heavily quantized versions of the model because their eval performance is terrible. Some providers advertise features on OpenRouter but will reject those requests when submitted to their API.
So pinning is the way to go.
For truly production use cases, use Novita, Fireworks, Toghether or something of the sort.
> Fireworks scored 46% on TAU, a 30 point gap
Another surprise was DigitalOcean being bottom of barrel too.
Companies are apparently willing to risk their brand name by being deceptive about these heavily quantized/flawed model-serving.
How many account credentials, balances, and tokens do you want to maintain? Even without automatic failover services such as openrouter are still incredibly useful.
Personally I pin a single vetted provider in the interest of minimizing risk.
I used to get content: "" all the time and I used to triple check my code to see if I was doing something wrong until I realized most AI providers in general have vibe coded their infrastructure as well and it is just a futile attempt to even fight it.
Providers probably serve quantized versions without disclosing it. Which is a real shame, because for certain tasks I would be perfectly willing to trade accuracy for cost. But, unfortunately, it is impossible to explicitly choose how quantized do you want your model to be, unless you are running it yourself on your own (or rented) hardware.
BTW, does anyone knows if LLM Gateway suffers from the same issues? Currently looking at trying it, but haven't got to it yet.
It should be OpenRouter's responsibility to protect you against it, by regularly benchmarking providers and giving you the control to avoid bad providers. In fact, that's a big opportunity for them, since it justifies their place as a middleman between users and inference providers.
If OpenRouter wants to succeed as a business, they need to be auditing the providers they connect to (i.e. benchmarking) and removing fraudulent ones from their service.
Source: Enterprise customer doing $XXM annual run rate of inference spend on their platform
Android app called "AudioRun"
https://play.google.com/store/apps/details?id=com.audiorun.a...
https://apkpure.com/audiorun/com.audiorun.app
If you for some reason had unusee balance and forget about it; its just gone.
Its a google model. The edge cases are crazy to deal with and have taken a long time to find. I am also finding that since it has no fallbacks but no published rate limit I am single handedly taking the model down on what I thought were a reasonable amount of request. There is no other fallback that isn't google. I'm worried that with sustained usage from my users on launch in a few days. Clearly google can't be making that much money on it it im like 10% of the usage and the top 2-3 user of it spending several hundred a month in pre launch testing.
So why would they care to help a small time start up? I will have to jump to a cost effective different model and find the footguns all over again at somepoint.
Not sure where im going with this but just wanted to let others know </rant>
Yay google, love living in fear of a vendor that they will rug pull me at any point because they can.
I think the rate limit issue happens because of OpenRouter sending so many requests to Google.
--
Update: oh, that is the AutoExacto Benchmarks! Now I see it.
If you’re building vision-native apps, there are so many footguns in vLLM/SGLang serving configurations, let alone the routing/orchestration in providers like OR, that leave the user more confused about the model’s capabilities.
Thanks everyone for the feedback here. Some of this we are aware of, some of it we aren't. Some we can fix, some of it is inherent to inference (and we in fact improve the situation dramatically).
Philosophically, at OpenRouter we are trying to do two different things, that are sometimes at odds with one another:
1. Let you use a lot of capacity across a lot of providers, in a way that "just works" and you don't need to worry about it.
2. Have a huge variety of inference available so you can pick radically different price/performance tradeoffs, data policy decisions, geographic destinations, inventive hardware, etc.
These are inherently odd bedfellows, and we are still very much improving how we can make both of them true at the same time.
Some quick thoughts on the article itself:
1. Benchmarks: YES! Providers benchmark differently. We run benchmarks on the live endpoints continuously, monitor the median performance, and kick providers out of the default routing pool if they vary by more than a standard deviation. We work hard (and continue to invest) to make sure that providers serving sub-par inference can't game the system, and that our routing actively avoids them. So the chart is accurate (it's our chart) and it actively influences our routing decisions!
2. That is bad and we will fix it. Sorry.
3. When we on-board providers we run essentially the same test as the author did to verify that the param is working as expected. If it isn't, we don't launch the provider. However this is not one we are running constantly in production. We are working on making this more robust in general and I do believe is fundamentally solvable in a way where it will "just work".
4. We 100% agree that users should not filter by quantization. It's a bit of a legacy concept in general; there is a huge amount of code between "model weights" and "inference API" and in almost all cases quality degrades in that part of the stack, NOT in the model weights themselves.
5. Hmm...we will dig in here. We monitor tool calls in real time and route around providers that are regularly mis-parsing tool calls. So you should get a very low rate of these in general. Another area we have invested a lot in: https://openrouter.ai/docs/guides/routing/auto-exacto
6. We will dig in here as well. I'm surprised this is happening frequently enough to be noticeable. We eat the cost when the finish reason is an error, but not when it is "stop". Perhaps we can expand our "insurance" program: https://openrouter.ai/docs/guides/features/zero-completion-i...
7. Will investigate.
8. We attempt to heal these, but obviously missed some. Will fix.
9. We do not rate limit by IP. Would love some more information here, as that is very surprising.
10. Ugh. That sucks. I'm sorry. We are introducing QoS tiers for production apps, which will address a lot of this.
(screenshot showing our internal testing of deepinfra image inputs: https://raw.githubusercontent.com/ping-Toven/images/main/ima... )
One thing to note about the first graph: nobody is doing as well as the first part on tool calling, and it's not close.
This might be the fault of the other providers, but it's probably just something slightly different that the first party does with the model inference program than anybody else, and that's not sure to weights it's due to vLLM twiddling (or whatever) and probably becuase the first part actually uses their own customized inference program rather than the standard methods that all the third party providers use. This isn't nefarious, it's just the challenge of these sorts of stochastic systems.
Having been in science for decades now, and seen benchmarking across many different fields, these results are completely expected for me. LLM serving is not mechanical, it's hard to get right and has lots of unknown footguns. Even something as extreme as scrambling a matrix will still likely get results that are nearly as good as normal, and if there's a bug deep in vLLM or the tensors metadata that results in that, then it's going to be pretty hard to find unless you're an active researcher with knowledge of the particular model you're running inference on. I kind of doubt that's happening here, but maybe!
In the scientific literature, when benchmarking methods, everybody's own method performs best in their own hands. Some attribute it to researchers gaming benchmarking for publication purposes, but I think it's just what we see here: the people who made a method are just the best at using it because they know all the quirks and use it best.
Programmers are not used to thinking with that nuance, and jump to conclusions about lying about quantizations, etc., but this is really just an unavoidable part of AI/ML methods: when things aren't perfect they're still pretty good and it's going to take the model creator to truly debug it. At least until the open weights ecosystem gets a lot better at ensuring reproducibility, and model cards are nowhere detailed enough for that to happen yet.
It's been a few months since I looked around at this topic, but Fireworks and Openrouter were the two options I (briefly) tried.
If you use any other meta-provider that routes your requests to third party providers, you'll likely face the same issues. If you try using any of those providers directly, you'll likely face some of the same issues as well, except you won't have the option of quickly swapping to a different one and taking your credits with you.
Extreme variance in quality and feature support per provider is probably the biggest obstacle holding back adoption of open weights models.
Not sure why people are drawn to this particular blunder. The promise of vendor neutrality maybe? I'll take working product over vendor-neutral slop anyways.
They used a closed list of 3 vendors in prioritized order, and got 429ed out of two of them, while the third one stopped serving the mode.
This is less of a problem if you're running an agent locally and routing your problem to OpenRouter - you can pin to one or two models for consistency and just switch models when something goes bad. But the article is specifically about production traffic.
Never before have I heard this sentiment, NGL. Vendor-neutrality has been an OS(/FLOSS) darling for, well, the whole time.
RE:"single vendor", if this post is to believed then you might have picked one that has 100x the tool calling errors for the next SoTA model, if your single vendor serves the next SoTA in the first place. It also completely erases the notion of competition driving down prices -- that would only hurt you in the short term, but obviously would ruin the whole ecosystem long term.
I feel like I must be missing something?
The most significant lock-in to me isn't even something you mentioned, but rather it's model related; I personally put a little time into trying to optimize my prompts every time I change models, as they all have their own unique... flavor.
As for tool calling errors, it seems like first party providers are among the best, I got the feeling that's what he was suggesting, though of course that's why you test. You can also go directly to Together.ai or whoever else you please.
Like others have said I think Openrouter seems neat for testing, but even just as a hobbyist I've been drawn to go direct to particular providers due to irritating little issues that I now see just weren't me.
Yesterday I wanted to do a bit of benchmarking a prompt across multiple models. Small requests. Outside the big providers, the experience became awful. This explains that experience.
It reads to me an argument for self hosting, maybe a less capable model to deal with smaller compute resources, but when starting to use that model and inference software to build a benchmark you can easily rerun when you change things to observe the impact of the change.
Maybe that’s a similar level of diligence but they feel different to me.
The worst case of hallucination I had, was DS v4 flash switching to Italian mid-conversation and impersonating a podcast host for no reason.
So I guess I shouldn't be surprised at all to see these benchmarks, but still I am!
There are such huge economies of scale with batched inference that it's clear this sort of service will continue, but it has a lot of growing up to do. Even AWS Bedrock has a Claude that feels different to me, but I haven't had a chance to do actual benchmarks that would show that.
They’re close enough to not matter though.
The models I've mainly been using recently are GLM 5.3 Flash and GLM 5.3. While obviously all these models have some variability, GLM 5.3 Flash feels like it oscillates between "I can't believe it's not Sonnet", but it costs a fraction of that and "This feels like I'm back using GPT-4, why am I even bothering with an LLM?".
https://newsletter.semianalysis.com/p/clustermax-20-the-indu...
This is actually expected behavior. No content and no tool call is the same as content-only: the agent decided it's done. Anthropic has done this for a while.
I wonder what tricks one could use to ensure the model is actually performing on par with the reference api, like matching seeds or running exact benchmarks.
"200 OK, no answer" - insane that openrouter's main feature is literally a fallback and streaming doesn't support 200 no content to fallback to another provider or smth.
"rate-limit by IP"... now it kinda makes sense why deepseek v4.1 flash rate limits me on prod but never seems to happen on local. Makes you have to basically pin Deepseek as provider, since I've never had 429 error on them
Is it, though?
https://app.answerhq.co/openrouter-ai/articles/credits/credi...
Noob question: do good harnesses automatically optimize for bad tool calling behavior automatically?
On top of that, near as I can tell, there are no protections for your API key. No restrictions by country, IP, etc...
I get it, the API key was my responsibility, but, setting the dollar limit is exactly the guard they suggest against that.
I was so disappointed that I won't consider any of them for at least a few years.
I feel like it's absolutely insane that some providers serve the same model with much less efficacy. That doesn't make sense to me. What's going on?! I'm suddenly feeling intense shame for having routed all my non-subscription usage through them so far, and honestly some white hot anger that they would blatantly lie about something so important.
What am I missing? Is this really true?
And errrr...yeah...that auto-exacto performance chart is both 100% useless, and totally unclear. We will get that fixed. But under the covers it is doing a lot of valuable work! https://openrouter.ai/docs/guides/routing/auto-exacto
It is pretty trivial to pin a single provider for a model. Better yet, instead of calling the model directly, use presets instead. You can easily change the setting on openrouter without having to update your app every time.
I feel there is still a lot of progress to be made before we can really trust LLM providers.
* Hardware availability
* Competency
* Scruples