Source: OpenAI — Rate limits guide

OpenAI's rate-limit guide: RPM/RPD/TPM/TPD/IPM per org and project, usage tiers (Free/Build/Launch/Grow), x-ratelimit-* headers, max_tokens reservation, ramp guidance and retry rules.

Contents

Summary

Explains the five rate-limit dimensions, how tiers are earned, how usage is counted, and how clients should react to 429/503.

Key claims

  • Dimensions: RPM, RPD, TPM, TPD and IPM (images); whichever is hit first throttles. Limits apply per organisation and project, vary by model, and some models share one pool. Long-context requests can have a separate limit.
  • Tiers (late 2026): Free, Build ($5 purchased), Launch ($100), Grow ($500), with monthly usage limits $100 / $500 / $5,000 / $200,000; upgrades are automatic. Example Grow limits: 15,000 RPM / 40M TPM for the Astra/Sol/Terra family.
  • Counted usage is the greater of max_tokens and a character-based token estimate, so an oversized max_tokens reserves capacity you never use.
  • Headers: Retry-After on temporary 429/503; x-ratelimit-{limit,remaining,reset}-{requests,tokens} plus project-scoped token variants.
  • Error variants: 429 for a temporary limit, 429 slow_down for rapid traffic increases, 503 server_is_overloaded for overload.
  • Failed requests count toward per-minute limits, so immediate resends do not help. Wait at least Retry-After plus jitter; otherwise exponential backoff with jitter; cap attempts and total retry time.
  • SDKs retry eligible 429 and 503 responses automatically, subject to their retry settings (the guide gives no default count); if you retry yourself, disable SDK retries to avoid multiplying requests. Never retry quota or billing errors.
  • Ramp guidance: past roughly 1M input TPM, grow by no more than 50% per 15 minutes.
  • The Batch API has its own queue limit (enqueued input tokens per model) and does not use synchronous RPM/TPM; a hard spend limit also returns 429.

Feeds

Rate limits and retries

Suggest changes →