Claude Opus 5.5 fast mode needs no access request, but falling back costs the cache

Anthropic has taken fast mode out of research preview on Claude Opus 5.5, so asking for it no longer needs an access request or a beta header. The release notes dated October 15 say the model "now has a default fast mode rate limit in the Start, Build, and Scale usage tiers", and that requests still sending the old fast-mode-2026-02-01 header keep working.

The Evaluation tier is left out until an organisation moves up. Anthropic's fast mode documentation says the new limit is separate from standard Opus rate limits, that going over it returns a 429 with a retry-after header, and that six response headers report how much fast input and output capacity is left in the minute.

Falling back to standard speed costs the prompt cache

The documented escape from a 429 has a bill attached. "Falling back from fast to standard speed will result in a prompt cache miss," Anthropic says. "Requests at different speeds do not share cached prefixes."

So an agent that catches the rate limit and retries at standard speed pays full input price for the whole prefix it had cached, which on a long coding context is the expensive part of the call. Anthropic's own sample code sets the retry count to zero on the fast attempt precisely so it fails straight away and falls back, rather than waiting out the two automatic retries the SDKs do by default.

Azure deployments get it, Anthropic-hosted Foundry does not

Fast mode on Opus 5.5 now runs on Claude in Microsoft Foundry for deployments hosted on Azure, and not for Foundry deployments hosted on Anthropic. Fast mode is also unavailable with the Batch API, unavailable with a Priority Tier commitment, and not currently available on Claude Platform on AWS.

That is the first time the feature has left the Claude API. When Opus 5.5 arrived on September 22 at $4 per million input tokens, fast mode was a research preview on the Claude API alone, and the premium is unchanged at $8 per million input tokens and $40 per million output, against $10 and $50 on Opus 5 and Opus 4.8.

Opus 5 and Opus 4.8 stay in research preview

Fast mode on both older models still needs the beta header and still needs permission, through an account manager or the waitlist Anthropic runs at claude.com/fast-mode. Opus 4.6, which had fast mode removed from it altogether, behaves differently again: a request that sets the speed to fast and sends the beta header succeeds and then runs at standard speed, with the response's own speed field reporting "standard" rather than returning an error.

None of this changes what the model does. Fast mode "runs the same model with a faster inference configuration", Anthropic says, with "no change to intelligence or capabilities". The claim is up to 2.5 times the output tokens per second, and the documentation is explicit that the gain sits in the output rate rather than the time to first token, so a streaming interface shows it and a short one-shot prompt barely will.

Anthropic publishes no tokens-per-minute figure for the new default limit and points at the Console instead, so how much fast capacity a Start-tier account actually has is something you find out by hitting the 429.

Note: Links to online stores in articles can be so-called affiliate links. If you buy through this link, MIXED receives a commission from the provider. For you the price does not change.