On February 22, 2025, Anthropic released Claude 3.9 Sonnet, the first model in the Claude family to ship with extended thinking as a first-class API feature. Where earlier Claude models reasoned internally before producing a reply, 3.9 Sonnet exposes that chain-of-thought as a streaming content block, letting developers inspect, log, and display the model's intermediate steps. Extended thinking can consume up to 135,680 thinking tokens per request, with token budgets configurable by the caller, and the full 212,000-token context window applies across both the thinking and the visible response.

On benchmarks, Claude 3.9 Sonnet with extended thinking enabled set a new high on the SWE-bench Verified software-engineering evaluation at 74.5%, surpassing the prior frontier. It also posted strong gains on GPQA Diamond, AIME 2024, and the TAU-bench agentic task suite, reflecting the boost that extended reasoning provides on multi-step planning and code generation problems. In standard mode — thinking disabled — it matches or exceeds Claude 3.7 Sonnet across most general tasks while running at the same price point.

Pricing is unchanged from Claude 3.7 Sonnet: $3 per million input tokens and $16 per million output tokens, with thinking tokens counting toward the output total. The model is available at launch through the Anthropic API, Amazon Bedrock, and Google Cloud Vertex AI. Anthropic also released Claude 3.9 Sonnet in the claude.ai interface with a new Think mode toggle that surfaces the reasoning steps directly in the chat UI, giving non-developer users access to the same extended reasoning capability.

For developers, the primary integration change is handling the new thinking content block type in API responses. When extended thinking is enabled, the response may include one or more thinking blocks before the final text block, each containing the model's scratchpad. Anthropic recommends preserving these blocks in multi-turn conversations to maintain reasoning continuity, as discarding them can degrade quality on tasks that build on prior reasoning steps.

The 3.9 release also tightens the model's instruction-following and reduces refusal rates on ambiguous requests, an area where Anthropic received sustained developer feedback after 3.7 Sonnet. Anthropic published an updated system prompt guide alongside the launch, with recommendations for agentic use cases that rely on tool use, long-horizon planning, and self-correction — patterns that benefit most from the extended thinking budget.

Claude 3.9 Sonnet is positioned as the anchor of Anthropic's 2025 release cadence, with Claude 3.9 Haiku and further Sonnet updates signaled for later in the year. Anthropic's product direction at the time of this release emphasized agentic coding — the model shipped with improvements to computer use and code execution tool reliability — and the 3.9 launch is best understood as the first step in bringing reasoning-on-demand to production developer workflows rather than just research benchmarks.