I looked into OpenRouter's free models and Free Router
Contents
I was curious about OpenRouter’s free models, so I dug into them. The short version: they are fine for trials and learning, but they are rough for production use.
what OpenRouter is
OpenRouter lets you call models from multiple LLM providers such as OpenAI, Anthropic, Google, and Meta through a single API. Pricing is passed through at the provider’s own rate, with no markup. The platform makes money through fees on credit purchases.
types of free models
selecting a specific model
You can choose a model by using the :free suffix in the model ID.
meta-llama/llama-4-maverick:free- 400B MoE, 17B active, multimodalgoogle/gemini-2.0-flash-exp:free- 1 million token contextmeta-llama/llama-3.3-70b-instruct:free- roughly GPT-4-class performanceqwen/qwen3-coder-480b-a35b-instruct:free- coding model aimed at agents
There are many others. You can check the free models collection.
the Free Router (openrouter/free)
If you send a request to the Free Router, OpenRouter randomly picks one available free model and runs inference on it. It filters models automatically based on the capabilities you need, such as vision or tool calling.
- context length: 200,000 tokens
- input: text + images
- output: text only
- cost: free
The downside is obvious: you do not know which model you will get, so output quality and style vary a lot. Debugging is also harder, and there is no way to keep using the same model throughout a session.
rate limits
| condition | limit |
|---|---|
| no credits purchased, or less than $10 | 50 requests/day |
| $10 or more in purchased credits | 1,000 requests/day |
| all users | 20 requests/minute |
Fifty requests per day is not much. It is fine for quick experiments, but if you plan to use it regularly, buying at least $10 in credits is the realistic path.
caveats for agent use
Some free models support tool calling.
qwen/qwen3-coder-480b-a35b-instruct:freeopenai/gpt-oss-20b:freearcee-ai/trinity-mini:free
That said, there is a known bug where tool calling against :free models can fail with No endpoints found that support tool use. Tool support may simply be limited in the free tier.
Running a serious agent workflow on the free tier is hard.
privacy note
With free models, your prompts and outputs are logged by the provider. Do not send personal or confidential information. Treat them as trial-only models.