Claude Sonnet 4.6

The Last Major Sonnet Baseline Before Claude 5

Sonnet 4.6 sits at an important architectural boundary.

It expanded the Sonnet tier to a 1,000,000-token context window and 128,000-token standard output limit while introducing adaptive thinking and general-purpose effort controls. At the same time, it preserved backward compatibility with the older manual budget_tokens extended-thinking API.

That makes Sonnet 4.6 different from both sides of the transition. Earlier Sonnet models rely on manual extended thinking. Sonnet 5 removes manual thinking budgets, changes tokenization, and turns adaptive thinking on by default. Sonnet 4.6 can use either adaptive thinking or deprecated extended thinking, and a request that omits thinking still runs without reasoning.

Anthropic launched the model with an emphasis on balanced speed and intelligence and specifically highlighted improved agentic search performance with fewer consumed tokens.

The model remains available as Active (legacy), but Anthropic recommends Claude Sonnet 5.5 for new deployments.

• Model ID: claude-sonnet-4-6.
• Released February 17, 2026.
• Retirement not sooner than February 17, 2027.
• 1M context window and 128K standard output.
• Text and image input with text output.
• Adaptive thinking supported; extended thinking deprecated but still accepted.

Search Quality Is a Trajectory Problem

Agentic search is different from answering from static context. The model may formulate a query, inspect results, decide whether evidence is sufficient, search again, fetch a document, filter irrelevant material, and repeat until it can synthesize an answer.

For this kind of workload, useful evaluation metrics include search count, evidence coverage, redundant retrieval, total token consumption, stopping behavior, source-grounded answer quality, and latency to a supported conclusion.

Sonnet 4.6 launched at the same time Anthropic made web search and programmatic tool calling generally available without beta headers, and added dynamic filtering for web search and web fetch.

That surrounding platform generation makes Sonnet 4.6 a useful historical baseline for search-heavy agents: it represents the mature pre-Claude-5 stack where tool orchestration, code execution, search, and large context were already central to production workflows.

A Large-Context Sonnet Before Claude 5

When Sonnet 4.6 launched, Anthropic described its 1M context window as beta. The model later moved to the full 1M context profile documented today.

That capacity makes it suitable for large repositories, long document collections, extensive conversation histories, tool results, retrieved evidence, and persistent agent state.

Standard output is capped at 128K tokens. Anthropic also documents up to 300K output tokens through Message Batches in beta.

The context window is large enough that poor context management can become an economic problem before it becomes a technical limit. Sending irrelevant state repeatedly costs money and can reduce the prominence of the instructions or evidence that matter most.

Prompt caching, retrieval, summarization, and selective history retention remain useful even when the model can technically accept a million tokens.

Stable Dateless ID and Legacy Reasoning Compatibility

The model ID claude-sonnet-4-6 is a pinned snapshot, not an evergreen alias.

Anthropic changed its versioning model starting with the 4.6 generation: a dateless ID identifies a fixed release rather than silently moving to newer weights.

Sonnet 4.6 also rejects assistant-message prefilling. This matters when migrating from Sonnet 4.5 or older applications that shaped output by pre-populating the final assistant turn.

Thinking-related sampling rules are less restrictive than on Claude 5, but thinking itself imposes constraints. When thinking is active, temperature must remain at its default value.

Sonnet 5.5 migration guidance also requires removing non-default temperature, top_p, and top_k values from older Sonnet integrations.

Finally, Sonnet 4.6 preserves older tool-orchestration patterns that later Sonnet models change, so migration testing should include the complete request body rather than just changing the model string.

A Broad Production Feature Set Before the Claude 5 Transition

The model accepts text and images and returns text. Adaptive thinking is the recommended reasoning path, while deprecated manual extended thinking remains available for legacy code. Effort can tune the overall amount of work the model performs even when thinking is not active.

Prompt caching lowers repeated-context cost. Batch processing cuts standard input/output rates by 50% and can expose up to 300K output tokens in beta.

Sonnet 4.6 is available across the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and Claude Platform on AWS.

Its primary limitation today is generation age: newer Sonnet models improve performance, lower token pricing, change reasoning defaults, and add more advanced orchestration controls.