Claude Sonnet 5.5: named by Anthropic, reportedly in partner testing, priced 89% by Sep 30

Anthropic has named Claude Sonnet 5.5 but has not dated it. Everything past the name is reported, not confirmed - and the one number that moves every day says it is close.

WHAT ANTHROPIC SAID

One sentence, in the Claude Opus 5.5 announcement on September 22: "Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks, with many of the same improvements to performance, efficiency, and safety." No model card, no id, no price, no date.

WHAT IS REPORTED

Posts on X between September 24 and 26 say Anthropic is stealth testing Claude Sonnet 5.5 under the id claude-sonnet-5-5, that some partners got access for 24 hours, and that a release could come around September 30 or October 1. The same posts list a 1M-token context window, 128K maximum output, adaptive thinking by default, forced tool use retired, and $2 input / $10 output per million tokens with cache reads at $0.20. None of it comes from Anthropic, and no outlet has independently verified the partner access. For scale: $2/$10 is exactly what Claude Sonnet 5 costs today, so the reported price would be no change.

WHAT THE MARKET SAYS

Polymarket's "Next Claude Sonnet released by...?" market at 10:47 UTC on September 27: 89% by September 30, on $8,263 traded, and 98% by October 31. The September 28 and 29 buckets trade thin ($1,603 and $948) and read 51% and 66%. The market resolves on the next Sonnet, whatever it is called. Its twin prices the next Claude Haiku at 92% by October 31.

WHAT OUR SWEEP SEES

Nothing yet. At 10:47 UTC on September 27 no Sonnet 5.5 or Haiku 5.x id appears in Anthropic's published model list, on OpenRouter, or in the models.dev catalog. Partner access, if it happened, would not show in any of them: they list what the public can call.

WHAT IS ARMED

The Sonnet hunt catches claude-sonnet-5-5 in every spelling a Claude launch has used: the bare id, a dated snapshot, OpenRouter's anthropic/claude-sonnet-5.5 and Bedrock's anthropic.claude-sonnet-5-5. It counts as a state-of-the-art launch, so it reaches everyone on the default setting.

Two fixes went in this weekend, and we would rather say so than have you find out on launch day. Until September 26 the Sonnet hunt sat outside the state-of-the-art scope, so its alert would have skipped subscribers on default settings. And our Haiku hunt was written for a Haiku 5 that never came: its pattern refused anything after the 5, so claude-haiku-5-5 matched nothing on the board. Both are fixed and pinned by tests.

The minute either model answers on a public API, the alert goes out.