OpenRouter free models in 2026, and their limits
Which OpenRouter models are free right now, how the :free suffix works, the daily request caps, and what free endpoints may do with your prompts.
On this page
OpenRouter lets you call some models for nothing. You don’t need a card or a trial, just a model ID that ends in :free. On 3 October 2026 its model API listed about 20 free text models, plus a router that picks among them for you.
The catch is a daily request cap and, for some of these models, your prompts.
What makes a model free on OpenRouter
A free model is its own entry with its own ID. Add :free to the ID of a model that has a free variant, and requests to it cost $0 per token. OpenRouter’s docs warn that a free variant “may have different rate limits or availability” than the paid one.
To browse them, open openrouter.ai/models?max_price=0 or drag the prompt pricing slider on the models page to Free. Ignore the image and video models in that view. Several show a $0 token price but bill per image or per second.
There’s also openrouter/free, a router added in February 2026. It sends each request to a free model of its choosing, which is handy when you don’t care which one answers.
The free models worth knowing
The free list leans on open-weight models from NVIDIA, Google, Qwen, Thinking Machines and a few smaller labs. GPT, Claude and Gemini have no free variants. These were the strongest picks for everyday use on 3 October 2026.
| Model | Lab | Context window | Good at |
|---|---|---|---|
| Inkling | Thinking Machines | 1M tokens | Reasoning, coding, images |
| Inkling Small | Thinking Machines | 1M tokens | The same family, smaller |
| Nemotron 3 Ultra | NVIDIA | 1M tokens | Reasoning and multi-step work |
| Nemotron 3 Super | NVIDIA | 262K tokens | Efficient multi-agent work |
| Gemma 4 31B | 262K tokens | Text and images, optional thinking mode | |
| Qwen3.8 27B | Qwen | 262K tokens | Coding, research, images |
| North Mini Code | Cohere | 256K tokens | Agentic coding |
The list rotates, and some models carry an end date. Two Poolside models on it today expire on 31 October 2026, so check the models page before you build on a free ID.
The limits
OpenRouter’s limits page sets two caps on every :free model.
- 20 requests per minute, for everyone.
- 50 requests per day until you’ve bought $10 of credits over the life of the account, then 1,000 per day.
Fifty a day is enough to try models and runs out fast in real use, because every message in a chat is a request. Failed attempts count toward the cap too, per OpenRouter’s pricing page, and extra accounts or keys don’t raise it.
If your balance ever goes negative, OpenRouter can return a 402 error even on free models, and they stay blocked until you add credits.
The 1,000-a-day tier takes a one-time $10 purchase, which comes to $10.80 by card once OpenRouter’s 5.5% fee ($0.80 minimum) is added. The $10 stays in your account as credits for paid models, though OpenRouter’s terms let it expire unused credits after a year.
What happens to your prompts
OpenRouter doesn’t store your prompts or responses unless you opt in, and its own logging program is off by default. The providers behind free models are another matter.
OpenRouter marks some providers, including NVIDIA, Thinking Machines and Liquid, as ones that may train on what you send. On 3 October about half the free models ran only on a provider marked that way. Your privacy settings have separate training switches for paid and free models, and per OpenRouter’s FAQ, it won’t route to a provider that logs unless the switch is on. If a free model never answers, check that setting before you blame an outage.
Treat free models as public. They’re fine for anything you’d type into a search box. Keep client work and anything near a password on a paid model whose provider policy you’ve read.
Free variant or paid version
Where a model is on both lists, the paid version wins on almost everything except price. Each free variant ran on exactly one provider on 3 October, while paid versions of popular models run on several, so OpenRouter can route around an outage. Paid variants have no platform-level request cap, and OpenRouter’s docs warn that free models may slow down at peak times.
Context windows can differ in either direction. Qwen3.8 27B offers 262K tokens free and 1M paid, while Nemotron 3 Ultra is the other way round. Read the model page rather than assuming.
When paying makes more sense
Cheap paid models cost very little. On 3 October, GLM 5.3 Flash listed at $0.15 per million input tokens and $0.50 per million output tokens, and GPT-6 Luna at $0.10 and $0.50. The $10 that lifts your free cap would buy 20 million output tokens at either rate.
Free models suit experiments and anything you’d happily post in public. For daily work, a cheap paid model drops the cap and the single provider, and most of the privacy question goes with them.
Using OpenRouter models in your browser
UtilEngine, our extension for Chrome, Brave and Edge, takes one OpenRouter key and puts Ask AI and Chat on every page. It doesn’t list :free variants, though. Its 85 models are paid endpoints, and the default, GLM 5.3 Flash, is one of the cheapest in that list. If you only want free models, use an app or a script that accepts any model ID.
If a few cents of usage is fine, setup takes a couple of minutes.
- Create an account at openrouter.ai and make a key under Keys. Give the key a spending cap.
- Buy credits. The minimum purchase is $5.
- In UtilEngine, open AI Settings and paste the key. UtilEngine checks it when you save, keeps it in that browser’s extension storage, and never syncs it.
- Pick a model per chat, and a different one for each AI command if you like.
UtilEngine costs $49 once, and OpenRouter bills your tokens at its own prices. Our bring-your-own-key guide and OpenRouter explainer cover the rest.
Common questions
Are OpenRouter’s free models really free?
Yes, $0 per token with no card needed. The costs are a cap of 50 requests a day until you’ve bought $10 of credits and, on some providers, your prompts.
What is openrouter/free?
A router that sends each request to a model from the current free list. It’s the quickest way to get a free answer, but you don’t pick the model, so quality varies from one request to the next.
Why did a free model I used stop working?
It may have left the free list or hit its end date. You may also have used up the day’s requests, or its only provider logs prompts and your free-model training switch is off. A negative balance blocks free models too.
Can I use OpenRouter free models in UtilEngine?
No. UtilEngine’s model list has paid endpoints only. Its default, GLM 5.3 Flash, is among the cheapest in that list.
Do free models have smaller context windows?
Sometimes. Several free models offer 1M tokens, but a free entry can be shorter than the paid version of the same model. The model page shows both.