Anthropic Releases Claude Sonnet 3.7 with Hybrid Reasoning & Adaptive Thinking Tokens

TL;DR & Executive Summary

• Anthropic launched Claude 3.7 Sonnet as the first unified frontier model combining standard high-speed inference and extended chain-of-thought reasoning within one model weight distribution.
• Developers can dynamically allocate a thinking budget from 0 to 64,000 tokens per API call, removing the friction of routing queries between two distinct models.
• Achieves a 93.2% score on SWE-bench Verified when paired with Claude Code CLI and Cursor agent environments, setting a new frontier for automated software engineering.
• API pricing matches Claude 3.5 Sonnet: $3.00 per million input tokens and $15.00 per million output tokens (including generated thinking tokens).

Architecture: The Hybrid Thinking Mechanism Explained

Until today, software engineering teams faced a difficult architectural tradeoff: route fast queries (such as boilerplate UI generation or concise chat responses) to low-latency models like Claude 3.5 Sonnet, or dispatch complex debugging, refactoring, and mathematical tasks to specialized reasoning models like OpenAI o3-mini. Claude 3.7 Sonnet solves this tradeoff by unifying both operating modalities into a single continuous weight distribution.

Through the Anthropic Messages API, developers pass a new parameter: thinking: { type: enabled, budget_tokens: 4000 }. When budget_tokens is set to 0, Sonnet 3.7 functions with sub-second time-to-first-token latency for instant completions. When allocated up to 64,000 tokens, the model invokes an internal scratchpad to simulate test cases, verify edge conditions, inspect multi-file dependencies, and correct logical errors before generating its final answer.

SWE-bench Verified & Developer Tooling Benchmarks

Standardized evaluations reveal major performance gains across autonomous coding benchmarks. On SWE-bench Verified, Claude 3.7 Sonnet scored 70.3% in standard mode and reached 93.2% when granted full thinking budget inside iterative scaffold environments like Claude Code and Cursor. In full-stack React and TypeScript refactoring tasks, the model demonstrated an 84% reduction in syntax hallucinations and unresolved module imports compared to Claude 3.5 Sonnet.

In mathematical problem solving, Claude 3.7 Sonnet scored 96.2% on the MATH-500 benchmark, outperforming standalone specialized reasoning models while maintaining the natural, concise conversational tone developers expect from Anthropic models.

Developer Implementation & Cost Optimization

To maximize efficiency in production, engineering teams can configure dynamic token routing based on task complexity. For standard conversational turns and user onboarding, set budget_tokens to 0. For multi-file refactoring, automated PR reviews, and complex SQL schema generation, configure a sliding budget of 2,000 to 8,000 thinking tokens.

• Prompt Caching Support: Input prompts retain 90% cost savings on recurring context headers and system instructions.
• Backward Compatibility: Fully drop-in compatible with existing Anthropic SDK client configurations by updating the model string.
• IDE Integration: Native support launched today in Cursor, Windsurf, Claude Code CLI, and Continue.dev.