Rate limits on free LLM APIs, explained

A rate limit is a published cap on how much you can send, often per minute or per day, measured in requests, tokens, or a vendor unit that is not a token.

Last updated 12 Nov 2026. The provider figures on this page follow the tracker snapshot of 12 Nov 2026, 08:59 IST.

What is a rate limit on a free LLM API?

A rate limit is a cap on how much you may use the API in a window of time. Providers commonly publish limits as requests per minute (RPM), requests per day (RPD), tokens per day, or a credit or vendor unit such as a monthly call allowance. The cap may also differ by model. When this tracker has a public figure, it appears in the “limits” field or in tokens per day. When we do not, we show Not published instead of guessing.

Limits exist to keep free tiers shareable and to protect the service. Hitting one is usually a signal to slow down, not a sign that your key is necessarily revoked.

What do RPM, RPD and tokens per day mean?

Requests per minute (RPM) is how many HTTP calls you may start in a minute, often counted at the project or key level. A chat request counts as one request whether the reply is short or long, unless the provider says otherwise.

Requests per day (RPD) is the same idea over a day. Some free plans publish a low daily request cap even when they do not publish a token cap. In our list, OpenRouter’s recorded free-model limit is an example of a daily request cap until credits are bought — always re-check their official page for the live number.

Tokens per day is a cap on the text you send and, usually, the text the model returns. A token is a small piece of text, typically a few characters or a short word, not a whole document. Providers do not all use the same tokenizer, so a token on one official page is not guaranteed to match another. If we have a numeric daily token figure, we store it in tokens_per_day and show it on the card. If we only have a prose allowance (“free experiment tier,” “10,000 Neurons/day”), we keep that wording and leave the numeric field empty.

Some official clouds use a unit that is not a token. Cloudflare Workers AI, in our list, publishes a daily Neuron allowance, and our notes say the token equivalent depends on the model. Do not convert Neurons into tokens unless the official docs give you a method.

Why is “Not published” so common in the limits column?

Because many providers describe a free tier without putting a single RPM or RPD number on a public page, or they hide the exact cap behind a login. We do not copy numbers from unofficial screenshots or from memory. A missing number is not “unlimited.” It may still be enforced.

The current snapshot therefore separates providers that publish a daily token number from those that do not. Sorting the whole internet by tokens per day would quietly drop useful official tiers that only publish a prose description.

Are all limits daily, and do they reset at midnight?

Not always, and not always in your timezone. Some windows are rolling 24 hours; some reset at midnight UTC; some are monthly; some are one-time signup credits that never refill. Our tokens_per_day field is only used when the public source is a daily token figure. A 1,000-calls-per-month trial, a 90-day credit, or a “while experimental” offer is stored in the allowance or limits text instead.

If you need to know the exact reset, the official rate-limit or billing page is the source. We will not invent a reset time.

What happens when I hit a limit?

Most compatible APIs return HTTP 429 (Too Many Requests) or an error body that mentions rate limits. Some SDKs raise an exception with the same status. The response may include a Retry-After header, or a JSON field with a wait time. If it does, wait at least that long. If it does not, back off — wait, retry once or twice, then stop. Tight retry loops can turn a daily cap into an outage for your own key.

A 401 or 403 is usually authentication or permission, not a rate limit. A 400 with a model error usually means the model id is wrong or not included in the free tier. Read the body before you retry.

How should I design a small app around a free cap?

Make one test request before you add a batch job. Log the model id and the status code, never the key. If you run an agent or an IDE plugin, set a local maximum number of calls, because tools may retry on their own. Prefer official providers when the job is important, because a relay can add a second, unpublished cap on top of the upstream one.

When you compare options, read the allowance as a sentence. “200,000 tokens/day on listed models,” “1,000 calls/month,” and “$5 starter credit” are not interchangeable. The first is a daily text cap, the second is a monthly request cap, the third is a credit that may end when it is spent or when a trial expires. Our best free LLM APIs tables keep those distinctions instead of producing a single score.

Do limits vary by model?

Often yes. A provider may publish one cap for a small model and a tighter cap for a larger one, or they may share one pool across models. Our notes say so when the public source said so — Groq’s entry, for example, notes that limits vary by model, and the numeric daily token field is the published figure we recorded for listed models. The official rate-limit page remains the live authority.

If you switch model ids, do not assume the old cap still applies.

How does this relate to cards and trials?

A published free limit can still sit behind a credit card, a phone check, or a short trial window. Cerebras, in our data, is recorded as a 1,000,000 tokens/day allowance during a $5 30-day trial with a card required. SambaNova’s free plan is recorded as requiring a card and a purchase after a mid-October change. Those details belong in the allowance and notes, not in a hidden multiplier. Confirm them on the official plans page before you design around the cap.

How can I stay inside a cap without a published number?

Use small prompts, cap max_tokens when the API supports it, and watch the provider dashboard if one exists. Avoid concurrent workers until you know a request succeeds. If you receive 429s immediately, stop and read the official docs. Do not try to “measure” an unpublished limit by hammering the API; that can look like abuse and can get a key disabled.

This site’s last snapshot of published limits was 12 Nov 2026, 08:59 IST. For how to form the request itself, return to How to use free OpenAI-compatible APIs.