Anthropic ships Claude 3.7 Sonnet and Claude Code

Anthropic released Claude 3.7 Sonnet, which it described as its first “hybrid reasoning model” — a single model that could answer instantly in a standard mode or switch to an “extended thinking” mode in which it worked through a problem step by step before responding, with the reasoning process visible to the user. Developers using the API could set a budget for how long the model should think, up to 128,000 tokens, letting them trade cost and latency against accuracy on a given task. Pricing stayed at the same $3-per-million-input/$15-per-million-output tokens as the previous Sonnet model, with thinking tokens billed as output.

Anthropic reported state-of-the-art results on SWE-bench Verified, a benchmark of real-world software-engineering tasks, and on TAU-bench, an agentic tool-use benchmark, positioning the model’s coding ability as its main selling point over rival releases from OpenAI and Google. The model shipped under Anthropic’s existing ASL-2 safety standard following the company’s routine pre-release evaluation process; it was not judged to require stronger safeguards.

Alongside the model, Anthropic launched Claude Code, its first dedicated coding tool, in a limited research preview. Unlike a chat interface or IDE plugin, Claude Code ran as an agent in the terminal: it could search a codebase, edit files across a project, run tests, and commit changes to version control with the developer supervising rather than typing each step. Anthropic said early testers had it complete in one pass tasks that would normally take upwards of 45 minutes of manual work.

Claude Code’s command-line, agentic design was widely imitated over the following year: OpenAI’s Codex CLI and Google’s Gemini CLI both followed with terminal-based coding agents of their own, and the category it opened became one of the more competitive fronts in AI product development through 2025 and into 2026.