OpenRouter gives you genuinely free access to a rotating set of roughly two dozen models behind one OpenAI-compatible API key. Any model whose ID ends in the free suffix costs zero per token, capped at about 20 requests per minute and a daily quota that starts around 50 requests and jumps to roughly 1,000 once you have ever bought 10 dollars of credits. That is enough to build and test real things for nothing, and not enough to run anything in production. Here is exactly what you get, why it exists, and when to stop relying on it.
For per-model paid prices across providers, see the rankings. For every no-cost key worth having, see free LLM API keys.
How free routes actually work
OpenRouter is a router, not a model host. It forwards your request to an underlying provider, and some providers choose to serve capacity at zero cost. Those endpoints show up in the catalog as separate model variants with a free suffix on the model ID. Three things follow from that:
- Free is a property of the endpoint, not the model. The same model often exists twice: a paid route with normal per-token pricing and a free route with rate limits. You pick which one you call by the model ID.
- There is an auto-router for free capacity. OpenRouter ships a meta-model, openrouter slash free, that automatically picks from whatever free models are currently available. Handy for throwaway scripts where you do not care which model answers.
- The list churns constantly. Models gain and lose free status week to week. In mid 2026 the free collection has hovered around 25 to 29 models, and popular ones have dropped off with no warning. Never hardcode a free model ID into anything you care about without a paid fallback.
The rate limits, plainly
OpenRouter documents the limits clearly, and they are the whole story of whether free routes work for you:
| Limit | Without credits | After buying 10 dollars of credits (ever) |
|---|---|---|
| Requests per minute on free models | 20 | 20 |
| Requests per day on free models | ~50 | ~1,000 |
| Per-token cost on free routes | 0 | 0 |
A few practical notes:
- The 10 dollar threshold is cumulative and one-time. You do not need to keep a balance topped up; having purchased 10 dollars of credits at any point unlocks the higher daily cap. It is the best 10 dollars a solo builder can spend.
- The 20 requests per minute cap does not increase with credits. Free routes are throughput-limited by design.
- Failed requests can still count against your daily quota, and free endpoints get congested at peak hours. Expect occasional slow responses and retries.
- Limits change. Treat the numbers above as the shape of the deal, and check OpenRouter's own limits page before you build around a specific figure.
Which model families show up free
The exact list moves weekly, but the free pool consistently draws from the open-weight world:
- Qwen, including coder variants with very large context windows. Frequently the strongest free coding option; see the Qwen API guide for the family.
- DeepSeek, whose reasoning and chat models rotate in and out of free status. DeepSeek is also cheap to call directly; see DeepSeek API free credits.
- OpenAI's GPT-OSS open-weight models, in small and large sizes, solid general-purpose picks.
- Llama, usually the mid-size instruct models.
- Mistral and NVIDIA Nemotron, plus a long tail of experimental and community models.
Frontier closed models, the ones at the top of the LLM API pricing comparison, essentially never appear free. The free pool is a showcase for open-weight models, and that is fine: the best free coding models now benchmark close to paid mid-tier models from a year ago.
Why anyone serves models for free
Nobody burns GPU hours out of charity. Free routes exist for two reasons, and you should know which one applies to your traffic:
- Marketing. Hosts and model labs use free capacity to get developers hooked on a model before it goes paid, or to demonstrate their serving stack.
- Training data. Some free endpoints reserve the right to log your prompts and outputs and train on them. OpenRouter surfaces each provider's data policy, and your account settings include toggles to refuse routing to providers that train on your data, with separate switches for free and paid models.
If you are sending anything sensitive, flip those toggles or skip free routes entirely. The price of zero is sometimes your data. That trade can be completely fine for tinkering and completely unacceptable for customer data. Decide deliberately.
When free routes are the right tool
- Prototyping and model shopping. One key, dozens of models, zero spend. This is the fastest way to compare open-weight models before committing.
- Personal tools and agents. A daily quota of around 1,000 requests covers a lot of personal automation, especially paired with other free tiers like Groq, Gemini, and Mistral.
- Evals and experiments where a slow or occasionally failing request costs you nothing but a retry.
When to jump to paid
Move off free routes the moment any of these is true:
- You need reliability. Free endpoints deprioritize you under load and can vanish overnight. Production traffic needs a paid route or a direct provider key.
- You need throughput. 20 requests per minute is a hard wall. Paid OpenRouter routes and direct APIs scale far beyond it.
- You need provider features. Batch endpoints and prompt caching, which can cut a real bill in half, live on paid and direct routes. See prompt caching and batch APIs.
- You care about data handling. Paid routes give you far more control, including zero data retention options.
When you do go paid, the next question is whether to stay on the router or call providers directly. That math depends on volume and is covered in OpenRouter vs going direct. Compare the actual per-token prices for your shortlisted models in the rankings, which we keep current precisely because these numbers move faster than blog posts.
Bottom line
OpenRouter's free models are the best zero-cost LLM playground available in 2026: one key, a rotating pool of capable open-weight models, and limits generous enough for real prototyping, especially after a one-time 10 dollar credit purchase raises your daily cap. Treat it as a lab, not infrastructure. Know that some free endpoints may train on your prompts, keep a paid fallback in your config, and graduate high-volume traffic to the cheapest paid route once something works. Perkstack tracks free tiers, credits, and live per-model prices in one place: browse the catalog and create an account to keep tabs on what is free this month.
Related reading: OpenRouter vs going direct, free LLM API keys, the cheapest way to run LLMs