Source: OpenAI — Fast mode (formerly priority processing)

OpenAI's pay-per-token premium tier (`service_tier: "fast"`/`"priority"`): faster, steadier latency, ramp-rate downgrades to Standard, shared rate limits, and its relation to Scale Tier.

Contents

Summary

Fast mode buys lower and more consistent latency per token, without a commitment, for user-facing traffic.

Key claims

  • Set service_tier to "fast" or "priority", or make Fast the project default. Up to 2.5x faster than Standard for gpt-5.6-sol, with more consistent latency. For GPT-5.6 and earlier the response reports priority whichever value you send.
  • The response’s service_tier field reports the tier actually used; a request downgraded by the ramp limit reports "default" and is billed at Standard rates.
  • Price premium: GPT-5.6 Sol at 2x Standard ($8/$40 per 1M input/output short context, $16/$60 long context, promotional through at least 2026-11-21).
  • Fast and Standard share the same per-model rate limit. Ramp rule of thumb: past 1M input TPM, add no more than 50% per 15 minutes; the ramp limit is shared across all customers. Shift traffic over hours with feature flags; don’t run ETL/batch jobs in Fast mode.
  • GPT-6 Astra Fast mode has no latency SLA; for GPT-5.6 and earlier, Fast mode and Scale Tier share SLA treatment with service credits for eligible Enterprise contracts.
  • Fast mode is billed separately from Scale Tier and does not draw on Scale Tier units; Scale Tier spillover does not go to Fast mode. Heavy ramp-limit hits suggest buying Scale Tier.
  • Not available for fine-tuned models or embeddings, nor with EU data residency for GPT-6 Astra and GPT-6 Sol.

Feeds

Rate limits and retries, Provisioned throughput

Suggest changes →