Fast mode (Claude API docs): generally available on Opus 5.5

Summary

On 2026-10-15 Anthropic took fast mode out of research preview for Claude Opus 5.5: no beta header or access request is needed, and the Start, Build, and Scale tiers get a default fast mode rate limit. Fast mode runs the same model with a faster inference setup for up to 2.5x higher output tokens per second, at double the price. The same release notes added a between_tools thinking capability flag to the Models API.

Key points

• Enable it per request with speed: "fast" on claude-opus-5-5. The response's usage.speed says whether fast or standard was used.
• Price (checked 2026-10-20): Opus 5.5 $8 / $40 per 1M input / output tokens; Opus 5 and Opus 4.8 $10 / $50. Caching and data-residency multipliers stack on top.
• Opus 5 and Opus 4.8 stay in research preview and still need the fast-mode-2026-02-01 beta header (requests without it return 400). Sending the header to Opus 5.5 still works.
• Fast mode has its own rate limit with anthropic-fast-* response headers. A 429 means waiting (SDKs retry twice) or falling back to standard speed, which misses the prompt cache because speeds do not share cached prefixes.
• Available on the Claude API (including Managed Agents) and, for Opus 5.5, on Microsoft Foundry deployments hosted on Azure. Not available with the Batch API or a Priority Tier commitment.
• The speed-up is in output tokens per second, not time to first token.
• Same release notes: GET /v1/models now reports capabilities.thinking.types.between_tools, so code can check which models accept thinking: {type: "between_tools"}.

Why it matters

Fast mode on Opus 5.5 is now a normal, self-serve option for latency-sensitive work such as interactive coding agents. The cache miss on fallback means a fast-to-standard fallback costs more than it looks.