What are the Gemini, Groq and OpenRouter free tier rate limits right now?

[xyzs996]

Short answer, as published on 2026-07-25: Groq is 30 requests/minute and 1,000/day, OpenRouter's :free models are 20/minute and 50/day until you have bought credits, and Gemini does not publish a single free-tier number at all — it publishes a tier structure and sends you to AI Studio for the figures live on your project.

That last one is why this question is hard to answer by searching: two of the three have a number, one has an architecture.

Groq

30 RPM and 1,000 requests/day is the figure for llama-3.3-70b-versatile, and it is per model, per organization — not one shared account budget. The same 30 RPM applies to llama-3.1-8b-instant, but its daily ceiling is 14,400, not 1,000. gpt-oss-120b and gpt-oss-20b also get 30 RPM / 1,000 per day; groq/compound gets 30 RPM / 250 per day. There are token ceilings underneath all of these (12K TPM / 100K TPD on the 70b), and on a long-context workload the token cap binds well before the request cap does.

Source: https://console.groq.com/docs/rate-limits

OpenRouter

Any model ID ending in :free is capped at 20 requests/minute regardless of account status. The daily cap is the part people get wrong: under 10 lifetime credits purchased it is 50 requests/day, and at 10 or more credits it becomes 1,000/day. So "the free tier" is two different tiers depending on whether you have ever paid anything. OpenRouter also governs capacity globally, so opening extra accounts does not multiply it.

Source: https://openrouter.ai/docs/api-reference/limits

Gemini

Rate limits apply per project, not per API key, and depend on the project's usage tier. Free is the entry tier and moving up requires enabling billing. Google publishes the structure and then directs you to AI Studio for the RPM, TPM and RPD actually active on your project — so any blog post quoting you one universal Gemini free-tier number is quoting a snapshot of someone else's project.

Source: https://ai.google.dev/gemini-api/docs/rate-limits

A few more, since they come up in the same search

• SiliconFlow — 1,000 RPM, per account rather than per key, each model counted separately.
• SambaNova — 20 RPM / 20 per day / 200,000 tokens per day, and the free tier applies exactly while no payment method is linked.
• Cohere — 20/minute on chat, plus a monthly ceiling of 1,000 API calls on a trial key.
• Fireworks — 10 RPM across the entire account with no payment method attached.
• Cloudflare Workers AI — 10,000 Neurons/day, reset at 00:00 UTC. Neurons are compute, not requests, so there is no fixed request count to quote.

What these numbers are not

They are published limits, not measured throughput. Nobody here ran a load test; these are what the provider's own documentation says, with the date it was last read. A provider can also enforce something tighter than it documents. If you need to know what your key does right now, the repo ships a browser-side key checker: https://xyzs996.github.io/free-llm-api/verify.html

Every provider, filterable: https://xyzs996.github.io/free-llm-api/

One thing that would help: reply with the provider you are on and the limit you actually hit. Documented-versus-enforced gaps are invisible from the outside, and a single concrete report is enough to make me go re-read the source page and date it again.

[Ziuluiziul]

Small datapoint that might help people comparing free-tier RPM/TPM beyond Groq/OpenRouter/Gemini:

I ran a live smoke against Nous Research free models (:free on their OpenAI-compatible inference API). Docs list 50 RPM / 500k TPM for free, and the response headers expose x-ratelimit-* so you can verify remaining quota without guessing.

Sample from 2026-09-05 on meituan/longcat-2.0:free: HTTP 200, remaining 49/50 req and ~499962/500000 tokens after one tiny completion.

Full smoke + cotas report (incl. host/base URL notes): https://gist.github.com/Ziuluiziul/296fa307f82eb950e49cb54af08301eb

Worth treating published limits as a ceiling and checking the headers after a real call — some free tiers look fine on paper until the first 429.