How to calculate tokens per minute

By

Calculate how many requests fit in a minute from an RPM limit, a TPM limit, and the tokens in each request. You type every number.

Free toolThis tutorial uses the Token Rate Calculator. It runs in your browser, so open it in another tab and follow along.

A provider can cap you twice. RPM is how many requests you may send in a minute. TPM is how many tokens those requests may use. The token rate calculator turns the numbers you type into the smaller of those two caps. It has no provider table. You type every limit.

The numbers stay in the URL. Nothing is uploaded.

Run the sample

  1. Open the token rate calculator.
  2. Click Load sample.
  3. Read the summary. Copy writes it.

The sample is RPM 500, TPM 300000, 1200 input tokens, 300 output tokens, and 1000 requests.

The summary is:

Tokens per request: 1,500
RPM ceiling: 500
TPM ceiling: 200
Allowed per minute: 200
Binding limit: TPM
Batch tokens: 1,500,000
Minutes for 1,000 requests: 5

The notes are:

This page adds input and output. It does not know which tokens a provider counts. If your provider counts only input tokens, put 0 in output.

A blank limit means that limit is not applied. This page does not model a fixed or sliding window.

Tokens per request are 1200 plus 300, which is 1500. The TPM ceiling is 300000 divided by 1500, which is 200. The RPM ceiling is 500. 200 is smaller, so TPM binds. 1000 requests at 200 per minute is 5 minutes. The batch is 1000 times 1500 tokens, which is 1,500,000.

Change the request count to 250. The last line becomes:

Minutes for 250 requests: 250/200. That is 1.25 at a steady rate

1.25 is the division. It is not a claim about when a provider’s window resets.

One limit, or a request that does not fit

Clear the TPM box and set RPM to 100, input tokens to 10, output tokens to 0, and requests to 250. The TPM ceiling is none. The binding limit is RPM. Allowed per minute is 100. The last line is:

Minutes for 250 requests: 250/100. That is 2.5 at a steady rate

Output is 0 because you said the provider counts only input tokens. If both token boxes are 0 and a TPM limit is set, the page stops. The error is: Tokens per request are 0. A TPM limit needs a token count. If your provider counts only input tokens, put 0 in output.

A blank RPM and a blank TPM is an error: Type an RPM limit, a TPM limit, or both.

A typed 0 is rejected. The error for RPM is: An RPM limit of 0 allows no requests.

If one request uses more tokens than the TPM limit, the TPM ceiling is 0. The note is: One request uses more tokens than the TPM limit, so no request fits in that minute. The batch line says the batch does not fit in one minute.

How this page turns RPM and TPM into a request rateInput tokens plus output tokens is the size of one request. The RPM ceiling and the TPM ceiling are compared. The smaller ceiling is how many requests fit in one minute.Input plus outputTokens per requestRPM limitRPM ceilingTPM divided bytokensTPM limitSmaller ceiling

MCP server design: tools, authentication, and rate limits is about putting a request limit and a spend limit on your own product. Vercel AI Gateway tutorial is one place a provider limit shows up as a 429. Background jobs in Node.js with BullMQ is where a retry waits after that 429.

Inference cost prices tokens. Token cost is the other cost page. Exponential backoff is the wait between retries, which is a different number from how many requests fit in a minute.

Try the Token Rate Calculator

Turn RPM and TPM limits into requests you can send per minute. You type the limits and the tokens per request. The page does not look up a provider rate card.

Open the Token Rate Calculator →

  • Inference Cost Calculator: Estimate your monthly LLM inference bill and compare models across OpenAI, Anthropic, Google, and more.
  • Token Counter & Prompt Cost: Paste a prompt, get an approximate token count and what it costs as input across every major LLM.
  • Exponential Backoff Calculator: List the wait before each retry from a base delay, a multiplier, and a cap. Compare no jitter with full, equal, and decorrelated jitter.

More tool tutorials →

Want me to talk about your product? You can sponsor this site.