Claude 3.7 Sonnet and Claude Code

Source: Anthropic Date: 2025-02-24 URL: https://www.anthropic.com/news/claude-3-7-sonnet

Summary

Anthropic launched Claude 3.7 Sonnet on February 24, 2025, as “the first hybrid reasoning model on the market” — toggling between instant responses and extended step-by-step thinking. Pricing unchanged: $3/M input, $15/M output. State-of-the-art on SWE-bench Verified and TAU-bench. Simultaneously introduced Claude Code as an agentic command-line tool (limited research preview). Early testers: Cursor, Cognition, Vercel, Replit, Canva. Claimed 45% reduction in unnecessary refusals vs. predecessor.

Implications

• Claude model cadence / Claude Code thread. Claude 3.7 Sonnet and Claude Code are launched together — the model and the product are co-announced, establishing the pattern that major capability improvements come with an agentic product release. This is the Claude Code origin event.
• “First hybrid reasoning model.” The extended thinking mode (switchable) predates o3’s comparable feature and is distinct from GPT-4 Turbo’s architecture. Anthropic claims the “first” — the technical claim is that it can switch vs. being fixed-mode reasoning.
• SWE-bench leadership. State-of-the-art on SWE-bench at launch positioned 3.7 Sonnet as the leading coding model for the period between February and August 2025 (when Opus 4.1 improved it to 74.5%). Cursor and Cognition as early testers reflect immediate developer tool adoption.
• Claude Code as limited preview. Launching Code as a “limited research preview” is intentionally constrained — Anthropic is testing the agentic product in a gated context before broad release, consistent with the minimal-footprint agent framework published the same period.
• Watch: how quickly Claude Code went from preview to GA; whether the hybrid reasoning mode drove measurable improvement in production use cases vs. pure benchmarks.