Claude Sonnet 5.5

What is disclosed

Claude Sonnet 5.5 is the second model of the Claude 5.5 family, released on 28 September 2026.¹¹1Anthropic, Introducing Claude Sonnet 5.5, 28 September 2026. https://www.anthropic.com/claude-sonnet-5-5 Like every current Claude model, it comes without a published size or architecture.

Its limits are those of Opus 5.5. The window holds 1M tokens and a response up to 128K, or 300K on the Message Batches API with a beta header.²²2Anthropic, Claude Sonnet 5.5 model page and What's new in Claude Sonnet 5.5, Claude Platform docs, read 29 September 2026. https://platform.claude.com/docs/en/models/sonnet-5-5/overview and https://platform.claude.com/docs/en/models/sonnet-5-5/whats-new-sonnet-5-5 It takes text and images and answers in text. What it knows reliably stops at June 2026.²²

Thinking is adaptive and on by default, at high effort.²² It cannot be disabled. The lowest setting, between_tools, turns off the thinking up front and works only at high effort or below.²² Any sampling setting other than the default returns an error.²²

What is new

Sonnet 5.5 costs what Sonnet 5 costs, $2 per million input tokens and $10 per million output tokens, with cache reads at $0.20.³³3Anthropic, Claude API pricing, read 29 September 2026: Sonnet 5.5 and Sonnet 5 both at $2 input, $10 output and $0.20 for cache reads; Opus 5.5 at $4 and $20. https://platform.claude.com/docs/en/about-claude/pricing. Arithmetic: 4 / 2 = 20 / 10 = 2. The saving is in tokens. Anthropic says the model needs far fewer for the same work and costs up to 30% less per task.¹¹ Output comes more than 30% faster.¹¹

On Terminal-Bench 4.0 it scores 70.6%, against 10.3% for Sonnet 5. That is above the 66.4% the same table gives Opus 5.5 at xhigh, and no effort level is stated for the Sonnet result.¹¹

FrontierCode 1.1 scores it lower at max effort than at xhigh, 46.2% against 52.1%. Sonnet 5 had 42.4%.¹¹ A footnote explains the drop. At max, the model more often ran a code review skill that splits the work across many subagents. In two cases this led to a timeout or to edits outside the task.¹¹ On GDPval-AA v2.1 it scores 1844 Elo, two points below Opus 5.5.¹¹

Because its cyber capability is comparable to Opus 5, it is the first Sonnet to launch with cyber safeguards.¹¹ Higher-risk security tasks fall back to Sonnet 5. The biology safeguards are Sonnet 5's, and they may flag some microbiology and virology requests in error.¹¹

Using it

The id is claude-sonnet-5-5.²² Anthropic sells it as a faster, cheaper complement to Claude Opus 5.5, which "remains clearly stronger" at complex, open-ended work.¹¹ Opus 5.5 costs twice as much per token.³³

Code that runs Sonnet 5 with thinking off has to switch to between_tools. A request that still sends thinking as disabled gets a 400 error.²² Forced tool use fails too, and thinking blocks are bound to the model and conversation that produced them.²²

Default effort depends on where the model runs. It is medium in Claude Code and the Claude apps, and high on the Claude Platform.¹¹ The Batch API takes 50% off input and output.²² Retirement comes no sooner than 28 September 2027.²²