Anthropic today released Claude 3.7 Sonnet, calling it the first hybrid reasoning model on the market. The model can produce near-instant responses or engage in extended, step-by-step thinking made visible to users through a scratch pad. API users can control the thinking budget up to 128,000 output tokens, trading speed for answer quality.

Alongside the model, Anthropic launched Claude Code, a command-line agentic coding tool in limited research preview. Claude Code can search and edit code, run tests, commit to GitHub, and execute terminal commands. The company says it has already become indispensable for their own team, completing some tasks in a single pass that previously took 45 minutes of manual work.

Claude 3.7 Sonnet achieves state-of-the-art scores on SWE-bench Verified (62.3%) and TAU-bench (81.2%), outperforming OpenAI’s o3-mini and o1 respectively on real-world coding and agent tasks. The model reduces unnecessary refusals by 45% compared to Claude 3.5 Sonnet, and is priced at $3 per million input tokens and $15 per million output tokens, consistent with its predecessor.

Extended thinking is available on paid Claude plans and API, while free users get the standard mode. The release comes as AI labs race to ship models; OpenAI is reportedly close to releasing its own hybrid model. In group chats this week, developers are debating whether a unified model is simpler than picking specialized models, and whether Claude Code signals the start of the agentic coding era.