Leaks put Claude Sonnet 5.5 at $2 input, release next week

Anthropic partners have received a second Claude Sonnet 5.5 checkpoint, better than the first, and the model is due out in the coming week, the leaker Lyra claimed on X on September 26. The $2 input price circulating with the leak is Lyra’s figure too, and TestingCatalog says it could easily change at launch. Lyra’s September 24 post gave the model a 1M-token context window; a Droid registry entry that LuminaBench shared on September 27 shows 872K.

When Anthropic launched Claude Opus 5.5, the first model in its Claude 5.5 family, it wrote that “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.” TestingCatalog says Sonnet 5.5 currently appears to be in random A/B tests on Claude and has been spotted on Factory, where it is not available yet. The tester Chetaslua posted an alleged side-by-side video of Sonnet 5.5 and OpenAI’s GPT-6 Sol on the same prompts, at what Chetaslua called the “same price point”.

What the $2 compares with

Lyra’s earlier post put Sonnet 5.5 at $2 for input, $10 for output and $0.20 per million tokens for cache reads, and said partners had access for only 24 hours. LuminaBench, who shared the registry entry, said its credits should not be read as pricing. Lyra’s input and output figures match the price Sonnet 5 launched at on June 30: an introductory rate set to run through August 31, with $3 and $15 given as the standard price after that.

Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, against Lyra’s $2 and $10 for Sonnet 5.5. Its cache reads, though, also cost $0.20 per million, and Anthropic says cache reads make up the majority of agentic and coding work costs. At Lyra’s figures, the two models would charge the same for that part of the bill.

Context, thinking and tools

Lyra’s 1M tokens of context and 128K of maximum output are the same pair Anthropic’s documentation lists for Sonnet 5. The registry entry’s 872K is less than Sonnet 5’s window; it shows the same 128K of output and “standard low -> max reasoning”.

Forced tool use is retired, Lyra also wrote, so pipelines that force a tool call would need reworking if that carries into the release.

The same post said thinking is adaptive by default, which is how Sonnet 5 works too, and can be disabled at low, medium and high effort but not at xhigh or max. If that holds, Sonnet 5.5 would follow Opus 5’s rule, where requests that disable thinking at xhigh or max return a 400 error, and keep an option Opus 5.5 lacks: on Opus 5.5, thinking is always on, and requests that disable it return a 400 at every effort level.