Anthropic announced Claude 3.7 Sonnet on February 24, 2025, describing it as the first hybrid reasoning model on the market. The model can provide near-instant answers or engage in extended, step-by-step "thinking" whose intermediate steps are visible to users. On the same day the company introduced Claude Code, a command-line, agentic coding tool that is available as a limited research preview.
Availability and pricing
Claude 3.7 Sonnet is available across all Claude plans (Free, Pro, Team, Enterprise), as well as on the Claude Developer Platform, Amazon Bedrock, and Google Cloud Vertex AI. Extended thinking mode is supported on all surfaces except the free Claude tier.
Pricing remains the same in both standard and extended thinking modes: $3 per million input tokens and $15 per million output tokens. Thinking tokens count toward output. Via the API, users can cap the model's "thinking" budget to at most N tokens up to the model's output limit of 128K tokens, allowing trade-offs between speed, cost and answer quality.
Design philosophy and modes
Anthropic positions Claude 3.7 Sonnet as a single unified model that combines ordinary LLM behavior with controllable, deeper reasoning. Users choose whether the model should answer immediately or self-reflect before replying. In standard mode, Claude 3.7 Sonnet is an upgrade to Claude 3.5 Sonnet; in extended thinking mode the model performs internal reflection that improves performance on math, physics, instruction-following, coding and other tasks. According to Anthropic, prompting works similarly across both modes.
The company says it optimized Sonnet less for competition-style math and CS problems and more for real-world tasks that reflect business usage of LLMs.
Performance and third-party evaluations
Early testing from multiple organizations reported strong improvements in coding and full-stack development:
- Cursor: again best-in-class on real-world coding tasks.
- Cognition: superior at planning code changes and handling full-stack updates.
- Vercel: exceptional precision on complex agent workflows.
- Replit: able to build sophisticated web apps and dashboards from scratch where other models encountered failures.
- Canva: frequently produced production-ready code with better design taste and significantly fewer errors.
Benchmarks highlighted by Anthropic include state-of-the-art results on SWE-bench Verified (which evaluates models on real-world software problems) and strong results on TAU-bench (which tests agents on complex tasks involving user and tool interactions). The company also reported improved performance in internal gameplay tests.
Anthropic details their "scaffolding" approach: their default evaluation uses minimal tooling (a bash tool and a file editing tool operating by string replacement, plus a planning tool). For a higher-compute submission they ran multiple parallel attempts, rejected patches that broke visible regression tests, and ranked candidates with a scoring model. On the subset of 489 SWE-bench tasks that run on Anthropic's infrastructure, Claude 3.7 Sonnet achieved 70.3% with that higher-compute scaffold; without the scaffold it reached 63.7% on the same subset. Anthropic notes 11 of the original 500 test cases were incompatible with their infrastructure and are reported separately.
Claude Code: agentic coding from the terminal
Claude Code is an agentic tool that can search and read code, edit files, write and run tests, commit and push to GitHub, and invoke command-line tools while keeping the developer informed at every step. It is offered in a limited research preview.
Anthropic reports that Claude Code has already become indispensable internally for tasks such as test-driven development, debugging complex issues and large-scale refactors. In early tests, Claude Code completed tasks in a single pass that otherwise would take 45+ minutes of manual work. Planned improvements include more reliable tool calls, support for long-running commands, better in-app rendering and expanded self-knowledge of Claude's capabilities.
The company's GitHub integration is now available on all Claude plans, enabling developers to connect repositories directly to Claude.
Responsible development and safety
Anthropic says it conducted extensive testing with external experts to evaluate Claude 3.7 Sonnet for security, safety and reliability. The company reports a 45% reduction in unnecessary refusals compared with the prior model, as Sonnet better distinguishes harmful from benign requests.
A system card accompanying the release provides detailed Responsible Scaling Policy evaluations, discusses prompt injection risks and how they assess and train Claude to resist and mitigate such vulnerabilities, and explores how reasoning models can offer safety benefits by making decision processes more understandable.
Looking ahead
Anthropic frames Claude 3.7 Sonnet and Claude Code as steps toward AI systems that augment human capabilities by reasoning deeply, operating autonomously and collaborating effectively. The company invites developers to join the Claude Code preview and contribute feedback that will shape future improvements. Additional technical details, appendices and the full system card are provided by Anthropic alongside the announcement.
