Claude 3.7 Sonnet

2025-02-24 · Publisher : Anthropic · Domain : Artificial Intelligence

Anthropic launches Claude 3.7 Sonnet, its first hybrid reasoning model: fast answers or visible extended thinking, with an adjustable budget in the API. Claude Code arrives in preview.

What changes

  • Hybrid model: a single model produces fast answers or step-by-step thinking made visible to the user.
  • Fine-grained control in the API: the user sets a thinking token budget, up to the 128K token output limit.
  • 45% reduction in unnecessary refusals compared with Claude 3.5 Sonnet, according to the vendor.
  • Launch of Claude Code in research preview (command-line coding agent) and GitHub integration on all Claude plans.
  • Improved computer use, with measures against prompt injection.

### Specifications - Output limit: 128K tokens (thinking included), according to the announcement. Context window of 200K tokens (secondary source, not included in the announcement). - Modalities: text and images as input, text as output. - Availability: Claude.ai (Free, Pro, Team, Enterprise), Anthropic API, Amazon Bedrock, Google Cloud Vertex AI. Extended thinking is available everywhere except on the free plan. - API identifier: claude-3-7-sonnet-20250219 (retired from the API on February 19, 2026). - Knowledge cutoff: not disclosed in the announcement. ### Pricing | Mode | Input (USD / M tokens) | Output (USD / M tokens) | |---|---|---| | Standard and extended thinking | 3 | 15 (thinking included) | Same price as previous models in the Sonnet lineup. ### Benchmarks | Benchmark | Score | |---|---| | SWE-bench Verified (pass@1, no scaffolding) | 63.7% | | SWE-bench Verified (high-compute scaffolding) | 70.3% | | TAU-bench | announced as state of the art, numeric score not included in the text | Scores published by the vendor. ### Key takeaways Extended reasoning becomes a setting of a single model rather than a separate model, with a thinking cost that can be controlled through a budget. The launch of Claude Code opens up agentic use of command-line coding.

Links

Tags : LLM, Anthropic, Claude, Sonnet, Raisonnement hybride, Claude Code