speed: "fast" is not supported on Opus 4.7

What changed under your feet

speed: "fast" errors on Opus 4.7. That much is loud and easy to find. What makes the parameter treacherous is that it fails in two different ways depending on the model, and only one of them is an error. If you're testing model versions in sequence during an upgrade, it's easy to misread either as a general deprecation rather than what it is.

Three behaviours for one parameter

Supported — Opus 4.8, Opus 5 and Opus 5.5 accept fast mode on the Claude API only, at their own premium per-token rates. Rejected — Opus 4.7 returns an error for the same request. Silently ignored — on Opus 4.6, Anthropic's pricing page says a fast-mode request runs at standard speed and is billed at standard rates. No error, no warning, and no speed-up: a latency budget built on fast mode simply isn't met, and nothing in the response says why.

That last case is the one worth auditing for. Check the speed field reported back in the response usage rather than assuming a request that succeeded ran fast.

Why 4.7 is the odd one out

Fast mode was pulled for the 4.7 release and returned on the Opus generations after it, which makes it one of the few capabilities in the lineup that isn't a straight line from older to newer. Treating capability as monotonic across generations is a reasonable default almost everywhere else; this and the cache-minimum table are the exceptions.

The practical fix

On Opus 4.7, drop the speed field — the request then runs at standard speed. On Opus 4.6, drop it too, since it buys nothing. If fast responses matter and you aren't locked to either version, move to a generation that supports fast mode; mid-migration, that is usually the simpler fix. Fast mode is not offered on Amazon Bedrock, Google Cloud, Microsoft Foundry or Claude Platform on AWS at all, and it cannot be combined with the Batch API.

Don't assume fast-mode pricing carries over

Where fast mode is available, it is priced as a premium over the same model's standard rate, so the absolute rate moves with the model: a cheaper base model carries a cheaper fast mode. Confirm the rate for whichever model you land on rather than reusing a budget built on another — the all-models pricing table lists every fast-mode rate in one place.