specification: API Commons Rate Limits
specificationVersion: '0.1'
schema: https://raw.githubusercontent.com/api-evangelist/interface-research/main/schema/api-commons.yml#/$defs/RateLimits
provider: OpenAI
providerId: openai
created: '2026-05-04'
modified: '2026-08-27'
generated: '2026-08-27'
method: searched
source: https://developers.openai.com/api/docs/guides/rate-limits
sources:
  - https://developers.openai.com/api/docs/guides/rate-limits
  - https://platform.openai.com/docs/guides/rate-limits
  - https://developers.openai.com/api/docs/guides/error-codes
supersedes: >-
  The 2026-05-04 bulk-sweep version of this file, which carried tiers and header
  names but no `limits[]` array at all — so every consumer reading limit_count
  saw zero while the artifact looked populated. It also asserted per-tier RPM/TPM
  figures for GPT-5 that the rate-limits guide does not publish; those are
  removed rather than carried forward, because OpenAI states model-level numbers
  on the models page per model and does not tabulate them by tier.
tags:
  - Rate Limiting
  - AI
description: >-
  OpenAI throttles on five dimensions at once — requests per minute, requests per
  day, tokens per minute, tokens per day and images per minute, plus audio
  minutes per minute for streaming audio models — and whichever is hit first
  returns 429. The numbers themselves are per MODEL and per usage TIER, and
  OpenAI publishes them in two different places: the monthly usage ceiling is
  tabulated by tier in the rate-limits guide (recorded below), while RPM/TPM are
  published per model on the models page and are not tabulated by tier anywhere,
  so they are not reproduced here rather than guessed at. Tier advancement is
  automatic on cumulative spend. The runtime signal is good: nine
  x-ratelimit-* headers plus Retry-After, including a separate project-token
  triple, so an agent can pace itself without ever reading the docs.
headers:
  limitRequests: x-ratelimit-limit-requests
  limitTokens: x-ratelimit-limit-tokens
  remainingRequests: x-ratelimit-remaining-requests
  remainingTokens: x-ratelimit-remaining-tokens
  resetRequests: x-ratelimit-reset-requests
  resetTokens: x-ratelimit-reset-tokens
  limitProjectTokens: x-ratelimit-limit-project-tokens
  remainingProjectTokens: x-ratelimit-remaining-project-tokens
  resetProjectTokens: x-ratelimit-reset-project-tokens
  retryAfter: Retry-After
responseCodes:
  throttled: 429
  slowDown: 503
metrics:
  - {id: RPM, name: Requests per minute}
  - {id: RPD, name: Requests per day}
  - {id: TPM, name: Tokens per minute}
  - {id: TPD, name: Tokens per day}
  - {id: IPM, name: Images per minute}
  - {id: APM, name: Audio minutes per minute, applies_to: streaming audio models}
limits:
  - name: Free tier monthly usage limit
    scope: organization
    tier: Free
    qualification: User in an allowed geography
    metric: usd_per_month
    limit: 100
    timeFrame: month
    currency: USD
  - name: Tier 1 monthly usage limit
    scope: organization
    tier: Tier 1
    qualification: $5 paid
    metric: usd_per_month
    limit: 100
    timeFrame: month
    currency: USD
  - name: Tier 2 monthly usage limit
    scope: organization
    tier: Tier 2
    qualification: $50 paid
    metric: usd_per_month
    limit: 500
    timeFrame: month
    currency: USD
  - name: Tier 3 monthly usage limit
    scope: organization
    tier: Tier 3
    qualification: $100 paid
    metric: usd_per_month
    limit: 1000
    timeFrame: month
    currency: USD
  - name: Tier 4 monthly usage limit
    scope: organization
    tier: Tier 4
    qualification: $250 paid
    metric: usd_per_month
    limit: 5000
    timeFrame: month
    currency: USD
  - name: Tier 5 monthly usage limit
    scope: organization
    tier: Tier 5
    qualification: $1,000 paid
    metric: usd_per_month
    limit: 200000
    timeFrame: month
    currency: USD
  - name: Realtime WebSocket connection duration
    scope: connection
    metric: minutes_per_connection
    limit: 60
    timeFrame: connection
    source: >-
      openapi/_original/openai-openapi-master.yml — the Realtime error code
      websocket_connection_limit_reached documents a 60-minute connection ceiling.
    note: A hard cap rather than a throttle; reconnect to continue.
tiers:
  - {name: Free, qualification: User in an allowed geography, monthly_usage_limit_usd: 100}
  - {name: Tier 1, qualification: $5 paid, monthly_usage_limit_usd: 100}
  - {name: Tier 2, qualification: $50 paid, monthly_usage_limit_usd: 500}
  - {name: Tier 3, qualification: $100 paid, monthly_usage_limit_usd: 1000}
  - {name: Tier 4, qualification: $250 paid, monthly_usage_limit_usd: 5000}
  - {name: Tier 5, qualification: $1,000 paid, monthly_usage_limit_usd: 200000}
policies:
  - name: Automatic tier advancement
    description: >-
      Tiers advance automatically once cumulative paid spend crosses the
      threshold. There is no application step.
  - name: Per-model limits
    description: >-
      Each model carries its own RPM/TPM ceiling, summarised on the models page
      and shown for your organization at platform.openai.com/account/limits.
      OpenAI does not publish a tier-by-model matrix, so no per-model numbers are
      asserted in this artifact.
  - name: Above Tier 5
    description: Custom limits beyond Tier 5 are arranged through OpenAI sales.
  - name: Backoff guidance
    description: >-
      Honor Retry-After and back off exponentially on 429 rate-limit responses.
      A 503 "Slow Down" is different — it asks for a sustained rate reduction
      held for 15 minutes, not a per-request retry.
  - name: 429 is overloaded
    description: >-
      Five distinct conditions return 429 and only one is retryable. Read
      error.code before retrying. See errors/openai-problem-types.yml.