It took me quite a while to come round to OpenRouter. Originally I didn't understand why anyone would put a proxy between them and an LLM, but it actually adds some quite significant value:
1. By far the lowest friction way to support and try out all the models.
2. They offer billing caps! Most model providers still don't do this [EDIT: maybe they do, see reply comment], but if you're going to run anything in public it's very useful to have hard limits so it doesn't cost you $1m overnight because someone started abusing it.
3. Their rankings are one of the more interesting signals for which models are popular, despite their flaws (most OpenAI and Anthropic users don't go via OpenRouter, it's currently not possible to tell the difference between many users switching v.s. one "whale" changing their preferred model)
Given how API costs are becoming meaningful for a lot of companies now, having a provider like OpenRouter to help measure your spend and easily experiment with and switch providers feels like a valuable service.
Another neat thing is, they publish hourly caching states for ALL model/provider combinations. I did some research on it to come up with a provider tiers list and found a bunch of open-source 3rd party hosts are simply trash tier https://dirac.run/posts/cache-hit-rates-agents
Thank you so much for this! I've been working on exactly this problem this week (which OpenRouter providers have the highest cache rate on average) because cache cost is sometimes half your cost: I'd much rather use a provider with more input caching with a more expensive/better LLM. Your results and lists seem more comprehensive than what I've done so far. Very helpful!
Good points. The easy experimentation factor is helpful for development, though I would gently encourage everyone to migrate to the 1st party APIs for pricing at scale.
OpenRouter is also a good place to find free LLM access with a catch: You should expect that any inputs and outputs are going into someone's training database. Clearly anyone who can pay should be using paid models with privacy protections, but the free models have been great for learning and experimenting. Especially for younger people learning API programming and LLMs who may not have access to a credit card or funds.
> You should expect that any inputs and outputs are going into someone's training database.
True enough, in theory; but what exactly are you imagining would be a useful-enough signal in the OpenRouter request+response stream, that any company would want their data as training material?
Even a single OpenRouter-API-key-identified subscriber's traffic, may consist of an mixture of traffic from multiple different sessions, under potentially multiple different end-users. (Where, if the subscriber is doing security correctly, then their OpenRouter key lives on a gateway rather than in a frontend app; and so the only IP address / UA / etc OpenRouter sees is that of the gateway itself.)
And the traffic stream may also invoke multiple models, and provide multiple different system prompts for those models; which, while marked in the traffic (i.e. conveyed as part of each request), makes the resulting data much less useful in aggregate, than if it were all training data for one model with one system prompt.
Plus, there are no RLHF signals in OpenRouter data. Even if OpenRouter wanted to build a general model-neutral framework for collecting RLHF-type data, it can't force subscriber apps to do the UI-level stuff necessary to collect it (i.e. the things ChatGPT/Claude do, with "thumbs-down" buttons, A/B tested responses, etc.) Analysis would have to rely on pure transcript-level user sentiment extraction.
I've wondered this too - exactly how are our inputs and outputs useful as training data? So I asked Gemini. Apparently using negative sentiment in user or llm responses can serve as RLHF, and the human prompts can also serve as useful data for what problems the llms need to be able to solve. There's also that smaller models can train on and improve from data from larger models but that's less relevant when not switching models in context.
It’s interesting all the focus on opt-out from training. Sometimes I worry there is an intentional focus on that so people don’t think about the other ways the company might be profiting off our data. Like I pay for Anthropic and they don’t train on that but are they selling my “anonymized” usage data in some other way?
From what I recall, these companies don't offer any option to opt out of your session transcript data being used (and sold!) for "regular" adtech targeting purposes.
> Clearly anyone who can pay should be using paid models with privacy protections
Clearly, anyone who needs privacy should be using models with privacy protections. Some people build open source and the models will get the code anyway.
Not completely free. Besides the rate limits, new sign-ups get 1000 credits (requests), which when used up, are gone forever. Only users with business accounts can get free refills.
I love their product and use them myself. But where's the value proposition for investors? Unless they get purchased by one of the large cloud providers, they will get pushed out of the market sooner or later.
What's the value proposition for the typical AWS startup to go with openrouter, if Amazon offers similar rates with direct integration into all their other offerings?
The only reason OpenRouter can exist at the moment is because we are in the wild-west phase of this technology, and lots of people and companies are exploring. In 5 years they will have to have transformed their business fundamentally, or go the way of the dinosaurs.
Billing caps are underrated! I don't understand why they aren't present everywhere. As an indie dev there are some services I'm really hesitant on trying by fear of getting an enormous bill for a mistake, this is even more true with vibe coding IMO.
I’m just not sure they have a moat or a long term play? I put $20 in and tried a few models. Then I went right to the model provider to put in $1000 and avoid the middleman tax. Now imagine a big corp spending millions on AI. That’s a lot of middleman tax.
The way how you manage the caps in OpenRouter is how every metered API provider should do it: keys have limits, and you can change the limits, and you set the limits to refill periodically, and you can create as many keys as you want.
> Long-running tasks like batch mode completions and agent sessions may incur overages beyond your project spend cap.
> Billing data processing times can be delayed in AI Studio, up to around 10 minutes. You may experience overages beyond your project cap if billing data hasn't processed before more charges are accrued.
Out of interest, why OpenRouter over a free option like Cloudflare’s AI gateway or another paid option like Vercel’s — any specific benefit to OpenRouter you’ve found, or just first you used that’s good enough?
I didn't know about these options either. I am using Cline: Cloudflare isn't an option but Vercel is. My spending is pretty low overall now that I'm using local models much more but good to know that there are cheaper alternatives to try or at least suggest to others.
Other features I've just noticed:
- configurable prompt injection protection using OWASP regex (https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_In...)
- configurable PIM protection for outbound prompts
- input/output logging
- "JSON healing" to auto-correct minor hallucinations
Lots of other stuff too. The business model seems pretty simple and the value-add features don't look particularly expensive or difficult to copy.
The biggest benefit is that it creates competition among models. If more people use open weight models or models from other providers, it’ll be harder to ban them. Which is what OpenAI and Anthropic will try to accomplish. OpenAI by lobbying the Trump administration for favorable treatment (see Brockman’s MAGA PAC donations), Anthropic by using religious leaders and nonprofits to push “safety” justifications for difficult regulations.
As someone who uses OpenRouter extensively (and wrote an unintentional adjacent PR piece a few days ago: https://news.ycombinator.com/item?id=48317294 ), it's definitely the best way to try out new models without fiddling with each providers distinct APIs which is becoming a recurring concern as of late.
That said, I don't understand the people who use something a full agentic backbone with expensive models like Claude Opus with OpenRouter because that 5% surcharge is meaningful at that level of cost instead of going with the source API providers. But people are clearly doing it, and it's pure revenue.
There is a lot of dumb token spend right now - tokenmaxing and such. Economic cost of token is not being evaluated carefully because there is fomo and no one wants to be left behind. But folks are waking up to it, and dumb token spending is not sustainable and will revert.
>> it's definitely the best way to try out new models without fiddling with each providers distinct APIs which is becoming a recurring concern as of late
Cursor only supports a single model (Kimi K2.5) not made by the Big 4 labs (OpenAI, Anthropic, Google, xAI). Cursor is actually extremely bad at wide model support.
I use Cursor with OpenRouter for some projects and it's great. Most of the time I just use Auto and let Cursor use its model or choose. If I run out of quota, or I'm not getting what I want, I switch off Auto and use OpenRouter to pick Opus, Codex, or whoever(all are available). Can continue the same context if you want, type "please continue" in the agent prompt, and on you go.
Cursor has limits even when using your own key. I was even cut off using a local model. I guess they use some sort of harness that requires non-local resources? I'm not sure I've actually tried to use Cursor in a fully-offline scenario yet. Cline works well enough and doesn't require any sign-up.
Is the Open in OpenRouter the same as in Open AI? I couldn’t find any repository or hosted code. Thought it'd be a open source, self hostable tool with a cloud offering but seems its just the latter?
I assumed they were open source but now that I checked they are not, they say "Open" because they route to third-party open models. Yikes. Another VC crap layer?
One thing that OpenRouter makes easy is the ability to manage API keys (mint new ones, expiry/limits per key, etc.) that I wish that other providers would make possible/easier.
So many use cases, like sharing AI/assisted features externally, with the ability to use those features but also limit the fallout if its shared / used for other purposes, without jumping through more fallible hoops like safeguards etc.
I was sort of hoping that they were bootstrapped or at least non-VC funded. I'm wary of them introducing consumer-unfriendly revenue-generating schemes.
I still don't get the value proposition: You rarely have to use all the models, you will likely end up with a few for your workflow but there is a way to use them/try all if you wanted to, neato.
Also one scary issue I had with OpenRouter in the early days, I think I saw somebody else's context and there were weird Chinese characters, haven't touched it since.
agreed, unless you need to use all models i'm sitting here wondering why orgs would want to introduce third party risk into their pipelines for marginal cost and time savings
Congrats to the OpenRouter team for securing this round of funding.
The 5% surcharge for their pricing model may not be palatable to enterprises. In fact, the OpenRouter team could be a pivotal part of the enterprise GenAI stack if they can allow configurable, pluggable endpoints for routing directly to enterprise vetted endpoints to 1P/3P LLM APIs. A couple of large companies I’ve worked so far kinda have this system in place, albeit the dev and maintenance cost and of setting up such an “LLM gateway” could be significantly reduced with OpenRouter. I feel that this is largely an ignored, forgotten part of operating GenAI apps at scale.
> The 5% surcharge for their pricing model may not be palatable to enterprises
Enterprises appear to be paying the API rates which are 10x (1000%) what are available to individuals, so I would not be confident they are sensitive to a 5% price change.
That said, the attraction of OpenRouter to enterprise customers should be that they save you >5% on average for a product <5% worse.
I think that OpenRouter will continue to be very popular while there lots of experimentation in the LLM space, and while the "current favorite" model continues to change between various frontier labs.
After things begin to settle down, we'll probably see a consolidation of both frontier and open-source models - and then OpenRouter will become less useful, because that 5% overhead is well worth it when you want to try 20 models from 10 labs, but harder to stomach when you only need 5 models from 2 providers, and each of those providers has its own API knobs that you can tune to make things even cheaper.
Why does a company with a seemingly health business model that is already churning profits and doesnt require large CapEx, taking losses to capture users, need to be raising this kind of capital?
i wonder if it's partially because it's not a unique business model and subject to yet another VC-subsidized race to the bottom on things like token prices
One thing I haven't seen mentioned here yet and really like about OpenRouter is their openrouter "meta" model, that automatically routes the prompt to an appropriately capable model. Saves me a ton of money on not routing everything through Opus, but not giving me bad results when I ask something more complex, which gets autorouted to Opus.
OpenRouter is our primary provider for evaluation data, and we've been really happy with them!
I'm sure they're experiencing growing pains, but a larger model selection (and faster releases for open weights models), would keep us from using other providers. For example, it took much longer than it should have to get Qwen 3.6 ~30B class models released (almost 2 weeks if I recall)
An amazing service. I use its 20+ free LLM options to allow completely free usage of LibreOffice AI extension with no signup https://librethinker.com .
I'm banned from using the free options. At some point they flagged my account as having engaged in model training against their ToS. This despite my account using around £15 worth of tokens over several months, nearly entirely through BYOK providers.
The handful of times I did try a free model is when I used their chat interface to quickly compare a few open weight models with a single prompt. That's the only usage I can think which could have triggered the block on my account. Even still, what's the point in have the simultaneous chat feature if using it veers so quickly into a ToS violation.
Their support is beyond useless in helping understand the situation. I don't think I managed to speak to anyone other than Tony Bot (or whatever it was named).
OpenRouter’s biggest value to me is reducing switching costs between models. The markup matters at scale, but for exploration and early-stage development, the convenience is hard to beat.
I think subscriptions are not going to last for serious users. Great to use them while we can, but AI does not fit the “power user subsidizes free/cheap users” model, nor the “support tens of thousands of customers from a small number of cheap servers” model. Everyone is a power user, and everything is computationally expensive.
Chatbot windows are a waste of time compared to API tools when trying to make stuff.
Subscribing to a vendor locks you in to sudden price swings that the big 3 are happy to do. The market needs lubrication for competition and provider routers offer that.
> ... with participation from NVentures (NVIDIA's venture capital arm), ServiceNow Ventures, MongoDB Ventures, Snowflake Ventures, Databricks Ventures ...
Are tech companies FOMOing so hard that they're now all running AI venture arms themselves instead of you know, developing their own products? Except for NVIDIA who needs to keep pumping the bubble I didn't expect the others.
Well, at least for them, investing into AI is actually developing their own product. The push to replace "Actually Indians" [1] with LLMs is huge because large Western companies want to save even the pittances they're paying Indian body shops.
I like OpenRouter - lets me test out new model quickly and easily. I would still need a good functioning mobile application for it.
I think they should go in this direction: they should make their own Model Agnostic versions of whatever functionalities other AI companies are making. Examples
1. personal chat app
2. the chat app working with their own implementation of memory
3. coding harnesses that are model agnostic
When I think of OpenRouter, I should think of "model agnostic LLM tools".
In what way? They're just an API customer like any other and charge a bit more on top. Providers would have to carve out their usage terms to not allow resell, which does nothing besides lose customers to competitors. If they all did that then you would tap on the FTC's shoulder and suggest they do their job.
1. By far the lowest friction way to support and try out all the models.
2. They offer billing caps! Most model providers still don't do this [EDIT: maybe they do, see reply comment], but if you're going to run anything in public it's very useful to have hard limits so it doesn't cost you $1m overnight because someone started abusing it.
3. Their rankings are one of the more interesting signals for which models are popular, despite their flaws (most OpenAI and Anthropic users don't go via OpenRouter, it's currently not possible to tell the difference between many users switching v.s. one "whale" changing their preferred model)
Given how API costs are becoming meaningful for a lot of companies now, having a provider like OpenRouter to help measure your spend and easily experiment with and switch providers feels like a valuable service.
OpenRouter is also a good place to find free LLM access with a catch: You should expect that any inputs and outputs are going into someone's training database. Clearly anyone who can pay should be using paid models with privacy protections, but the free models have been great for learning and experimenting. Especially for younger people learning API programming and LLMs who may not have access to a credit card or funds.
True enough, in theory; but what exactly are you imagining would be a useful-enough signal in the OpenRouter request+response stream, that any company would want their data as training material?
Even a single OpenRouter-API-key-identified subscriber's traffic, may consist of an mixture of traffic from multiple different sessions, under potentially multiple different end-users. (Where, if the subscriber is doing security correctly, then their OpenRouter key lives on a gateway rather than in a frontend app; and so the only IP address / UA / etc OpenRouter sees is that of the gateway itself.)
And the traffic stream may also invoke multiple models, and provide multiple different system prompts for those models; which, while marked in the traffic (i.e. conveyed as part of each request), makes the resulting data much less useful in aggregate, than if it were all training data for one model with one system prompt.
Plus, there are no RLHF signals in OpenRouter data. Even if OpenRouter wanted to build a general model-neutral framework for collecting RLHF-type data, it can't force subscriber apps to do the UI-level stuff necessary to collect it (i.e. the things ChatGPT/Claude do, with "thumbs-down" buttons, A/B tested responses, etc.) Analysis would have to rely on pure transcript-level user sentiment extraction.
Clearly, anyone who needs privacy should be using models with privacy protections. Some people build open source and the models will get the code anyway.
What's the value proposition for the typical AWS startup to go with openrouter, if Amazon offers similar rates with direct integration into all their other offerings?
The only reason OpenRouter can exist at the moment is because we are in the wild-west phase of this technology, and lots of people and companies are exploring. In 5 years they will have to have transformed their business fundamentally, or go the way of the dinosaurs.
Everything has a cost of some sort. It's just who you're going to pay and what the currency is.
https://news.ycombinator.com/item?id=48319827
Anthropic: https://support.claude.com/en/articles/8977456-how-do-i-pay-... - you can pre-pay and get a hard cutoff.
OpenAI: https://community.openai.com/t/how-to-set-billing-limits-and... - last time I looked OpenAI had a soft but not hard limit, I guess they fixed that last year.
I remember bugging them both about this last year, I need to update my mental model!
Deepseek has a prepaid model. (Pretty impressive, what fits into 10 Dollar)
> Billing data processing times can be delayed in AI Studio, up to around 10 minutes. You may experience overages beyond your project cap if billing data hasn't processed before more charges are accrued.
https://ai.google.dev/gemini-api/docs/billing#project-spend-...
That's a soft cap, not a hard cap
Looks like Vercel even have their own leaderboard: https://vercel.com/ai-gateway/leaderboards/models
Surprising that they have Opus 4.8 and 4.6 listed on the leaderboard but not Opus 4.7.
Other features I've just noticed: - configurable prompt injection protection using OWASP regex (https://cheatsheetseries.owasp.org/cheatsheets/LLM_Prompt_In...) - configurable PIM protection for outbound prompts - input/output logging - "JSON healing" to auto-correct minor hallucinations
Lots of other stuff too. The business model seems pretty simple and the value-add features don't look particularly expensive or difficult to copy.
They don't list themselves on https://openrouter.ai/providers
https://openrouter.ai/openrouter/owl-alpha
coffee farmers -> middle man -> you
That said, I don't understand the people who use something a full agentic backbone with expensive models like Claude Opus with OpenRouter because that 5% surcharge is meaningful at that level of cost instead of going with the source API providers. But people are clearly doing it, and it's pure revenue.
Why not... Cursor?
OpenCode is much better at it.
its just a proxy
So many use cases, like sharing AI/assisted features externally, with the ability to use those features but also limit the fallout if its shared / used for other purposes, without jumping through more fallible hoops like safeguards etc.
Also one scary issue I had with OpenRouter in the early days, I think I saw somebody else's context and there were weird Chinese characters, haven't touched it since.
DAMN!!
That's 41+ million tokens every second. That scale is crazy for such a small team of 48-50 people overall.
Enterprises appear to be paying the API rates which are 10x (1000%) what are available to individuals, so I would not be confident they are sensitive to a 5% price change.
That said, the attraction of OpenRouter to enterprise customers should be that they save you >5% on average for a product <5% worse.
After things begin to settle down, we'll probably see a consolidation of both frontier and open-source models - and then OpenRouter will become less useful, because that 5% overhead is well worth it when you want to try 20 models from 10 labs, but harder to stomach when you only need 5 models from 2 providers, and each of those providers has its own API knobs that you can tune to make things even cheaper.
I'm sure they're experiencing growing pains, but a larger model selection (and faster releases for open weights models), would keep us from using other providers. For example, it took much longer than it should have to get Qwen 3.6 ~30B class models released (almost 2 weeks if I recall)
The handful of times I did try a free model is when I used their chat interface to quickly compare a few open weight models with a single prompt. That's the only usage I can think which could have triggered the block on my account. Even still, what's the point in have the simultaneous chat feature if using it veers so quickly into a ToS violation.
Their support is beyond useless in helping understand the situation. I don't think I managed to speak to anyone other than Tony Bot (or whatever it was named).
Edit:
Total usage over 1 year:
Claude Sonnet 4.6 $8.80
Gemini 3.1 Pro Preview $6.71
Claude Opus 4 $6.19
Claude Opus 4.1 $7.49
Gemini 2.5 Pro $10.06
Claude Sonnet 4.5 $12.74
GPT-5 Codex $2.56
Grok 4 $4.39
Gemini 2.5 Flash Image Preview (Nano Banana) $1.88
GPT-5 $7.30
Others $7.99
Subscribing to a vendor locks you in to sudden price swings that the big 3 are happy to do. The market needs lubrication for competition and provider routers offer that.
Are tech companies FOMOing so hard that they're now all running AI venture arms themselves instead of you know, developing their own products? Except for NVIDIA who needs to keep pumping the bubble I didn't expect the others.
Well, at least for them, investing into AI is actually developing their own product. The push to replace "Actually Indians" [1] with LLMs is huge because large Western companies want to save even the pittances they're paying Indian body shops.
[1] for those OOTL: https://www.reddit.com/r/ProgrammerHumor/comments/1l3rpow/ac...
I think they should go in this direction: they should make their own Model Agnostic versions of whatever functionalities other AI companies are making. Examples
1. personal chat app
2. the chat app working with their own implementation of memory
3. coding harnesses that are model agnostic
When I think of OpenRouter, I should think of "model agnostic LLM tools".