OpenAI tokens-per-minute (TPM) rate limit
Rate limit reached … tokens per min (TPM)
You’re pushing more tokens per minute than your tier allows, usually from large prompts or long contexts, not from too many requests. OpenAI returns 429 until the per-minute token window resets.
What causes it
- Large prompts or long conversation histories: token volume, not request count, blows the TPM limit.
- Streaming many long completions in parallel.
- Stuffing entire files or codebases into context on every call (common in coding agents).
How to fix it
- Trim context: send only the messages and files the model actually needs for this turn.
- Summarize or compact long histories instead of resending the full transcript every request.
- Spread large jobs over time, and back off + retry on 429.
- Move up a usage tier for a higher TPM ceiling.
Running an agent? Coding and research agents hit TPM limits fast because they re-send big contexts on every step. Trimming context in the agent and a provider without a per-minute token quota both help here.
Your tier's limits
OpenAI rate limits by tier →
Requests and tokens per minute for each usage tier, and how to move up one.
Keep working on a fixed monthly plan
Standard Compute has no tokens-per-minute cap. Each key has a request-per-minute ceiling (300 on Starter up to 2,000 on Pro) and a fixed monthly compute budget, so token-heavy agent steps are bounded by the budget, not by a per-minute token quota.
Plans from $19 a month with $19 of compute. Eligible new accounts can try $0.25 of compute first, no card. Trial terms
FAQ
What’s the difference between RPM and TPM limits?
RPM caps how many requests per minute you can send; TPM caps how many tokens per minute. Big prompts hit TPM first; many small calls hit RPM first.