Claude Opus 5.5 fast mode leaves research preview — no waitlist

What happened

On October 15, Anthropic took Opus 5.5 fast mode out of research preview. The release notes say access no longer needs to be requested. The fast-mode docs describe the same Opus 5.5 weights at up to 2.5× higher output tokens per second, at premium pricing.

Key facts

What it is: Faster inference on the same Opus 5.5 model — not a different checkpoint. Gains target output tokens per second, not time to first token.

How to call it: Set speed: "fast" on Messages API requests to claude-opus-5-5. No fast-mode-2026-02-01 beta header required for 5.5; old headers still work.

Pricing: $8 / $40 per million input/output tokens for Opus 5.5 fast mode (standard Opus 5.5 remains $4 / $20).

Availability now: Claude API (including Managed Agents) and Claude in Microsoft Foundry for Azure-hosted deployments. Anthropic-hosted Microsoft Foundry deployments are not supported. Not available: Batch API, Priority Tier commitments, and Claude Platform on AWS (as of the Oct 15 docs). Fast mode for Opus 5 and Opus 4.8 stays in research preview and still requires the beta header.

Rate limits: Start, Build, and Scale tiers get a default fast-mode rate limit. Evaluation-tier orgs need to move up a tier first. Responses report usage.speed as "fast" or "standard". Caching caveat: Switching between fast and standard invalidates the prompt cache; the two speeds do not share cached prefixes.

How to use it

Agentic coding burns output tokens. Paying roughly 2× for up to 2.5× OTPS is a clear trade when a human is waiting on a refactor, a PR review loop, or a live demo. Opening the gate without a waitlist means most API customers can A/B the cost in a day instead of filing for preview access.

Read the fast mode docs and confirm your org is on Start, Build, or Scale. Add "speed": "fast" (or the SDK equivalent) on latency-sensitive Opus 5.5 calls; leave standard speed for bulk or overnight jobs. Monitor the fast-mode rate-limit headers and usage.speed. Successful Opus 5.5 fast requests report "fast"; capacity failures return errors. Implement an explicit retry at standard speed if you want fallback, accounting for the cache miss. Scope each run with Define done for a small deliverable, score a few timed trials with Build a demo observation scorecard, and log any regressions with Write a reproducible bug report.