Claude Sonnet 5.5

Overview. Claude Sonnet 5.5 is Anthropic's Sonnet-class chat model, released on 28 September 2026 as a direct upgrade to Claude Sonnet 5. It is closed-weight and reasoning-first: adaptive thinking runs on every request, and Anthropic's effort parameter (max, xhigh, high, medium, low; default high on the Claude API) controls how much of it runs. Developers reach for it when the work is long — a 1,000,000-token context window, text and image input, and up to 128,000 tokens of synchronous output on the Messages API.

Key strengths. The defining capability is long input: a 1,000,000-token context window paired with a 128,000-token synchronous output limit, taking text and images in and returning text. Reasoning is native rather than optional, and Anthropic states that Sonnet 5.5's effort levels are recalibrated relative to Sonnet 5, so thinking volume is tunable per request rather than fixed. Published accuracy evidence is less settled than the positioning: vendor-quoted Terminal-Bench 4.0 results and third-party readings of the same benchmark diverge widely because they measure different effort configurations, so any single headline number should be treated as configuration-specific.

When to use it. Use this for long-document work — reviewing or synthesising large corpora, sustained multi-turn conversations, and code or research tasks where fitting a very large input into one request matters more than raw speed. The five effort levels let you set reasoning depth per call, running low or medium for straightforward generation and routine tool loops and reserving high and above for hard multi-step problems; Anthropic supports changing effort mid-conversation through a beta header. Text and image input makes it a fit for workflows built around screenshots, diagrams, and document pages alongside prose.

Limitations. Reasoning cannot be turned off — Anthropic returns a 400 for a disabled thinking setting — and a non-default temperature, top_p, or top_k is rejected the same way, so this is not a model for sampling-parameter tuning. Forced tool use is unavailable: only the default `auto` and `none` are accepted, which constrains harnesses that depend on guaranteed tool invocation. Output is capped at 128,000 tokens on the synchronous API, Anthropic publishes no parameter count or architecture details, and there is no open-weight or self-hosted option.

Sovereignty note. Anthropic is a US company, so this model sits under US jurisdiction with CLOUD Act exposure, and the weights are proprietary — inference runs on vendor-controlled infrastructure rather than yours. Anthropic's commercial terms state that inputs and outputs from its commercial products, including the Claude API, are not used to train its models by default, and zero-data-retention arrangements are available for eligible Claude API customers by agreement, subject to Anthropic's approval. EU residency is still not on offer, however: Anthropic's data-residency controls provide only `us` and `global` inference with US-only workspace storage, so it should not be treated as a Sovereign option until zero-data-retention or residency guarantees are confirmed for the specific endpoint in use.