Claude 3.7 Sonnet and Claude Code are announced

Background

By early 2025 AI models were widely used to help write software, mostly through chat windows or editor plug-ins that suggested code for a person to accept. An agentic tool goes further. It takes a goal, decides which steps to take, and carries them out with access to files, commands and version control. Measuring such tools is also difficult. SWE-bench Verified is a benchmark that tests whether a model can resolve real software issues.

A hybrid model lets one system switch between quick answers and longer reasoning, so a developer does not need a separate model for each kind of task. Running in the terminal also means the agent works in the same environment a developer uses for builds, tests and version control.

What happened

On 24 February 2025 Anthropic announced Claude 3.7 Sonnet and Claude Code. Anthropic called Claude 3.7 Sonnet the first hybrid reasoning model on the market.^[1]

Claude Code launched as a limited research preview. It is an agentic coding tool that runs in the terminal, and developers can delegate tasks such as searching and editing code, running tests and committing to GitHub.^[1] The same announcement made Claude's GitHub integration available on all plans.^[1]

A research preview is an early version offered to a limited group for feedback. Features and terms can change before a general release, so the February launch was an early stage rather than a finished product.

Key details

Claude 3.7 Sonnet can give near-instant answers, or use extended, step-by-step thinking that the user can see. API users can set a thinking budget of up to 128K output tokens, which limits how much the model may spend on reasoning.^[1]

On SWE-bench Verified, Claude 3.7 Sonnet scored 63.7% on a 489-problem subset without special scaffolding, and 70.3% with high-compute scaffolding. Scaffolding means the surrounding software that helps a model plan, call tools and retry.^[1]

Pricing stayed at $3 per million input tokens and $15 per million output tokens, with thinking tokens counted as output.^[1] The model was available on all Claude plans (Free, Pro, Team and Enterprise), on Anthropic's developer platform, on Amazon Bedrock and on Google Cloud's Vertex AI. Extended thinking was available everywhere except the free tier.^[1]

Impact

On 22 May 2025, alongside the Claude 4 launch, Anthropic made Claude Code generally available. It added beta extensions for VS Code and JetBrains, Claude Code on GitHub Actions in beta, and an SDK for building custom agents.^[2]

The February announcement combined three elements: a model whose reasoning depth can be adjusted, an agent that works in the developer's terminal, and integration with GitHub. The benchmark figures are Anthropic's reports on one test, and results on any particular codebase depend on the task.

The benchmark numbers and product descriptions above are Anthropic's own. Readers who want to compare them with other tools can check the announcements listed in the sources.