Claude 3.7 Sonnet is Anthropic's first reasoning model, released February 24, 2025. Its framing was deliberately contrarian: rather than a separate reasoning line like OpenAI's o-series, 3.7 is one model with a dial, answering instantly by default or thinking in a visible extended-thinking mode with a user-controlled token budget, reflecting Anthropic's stated view that reasoning is a capability, not a product category.1
What was new
Extended thinking applied verifiable-reward-style training atop the Claude stack, with the chain of thought shown to users; Anthropic published research on the faithfulness limits of those visible traces, feeding the monitoring debate on this wiki's alignment pages. The release also shipped Claude Code as a research preview, a terminal-native agentic coding tool that became a product line of its own, and reported state-of-the-art agentic coding results on SWE-bench Verified, as reported by the developer.12
Behavior
3.7 tuned refusals further toward nuance, with the system card documenting a large reduction in unnecessary refusals. Its coding strength made it the successor default in the developer tooling built on Claude 3.5 Sonnet, while rivals from Google DeepMind's thinking-mode Gemini tiers to DeepSeek's open R1 line defined the crowded reasoning field it entered, all benchmarked on the MMLU-successor and competition-math suites with the usual developer attribution.2
Safety
Shipped at ASL-2 under the Responsible Scaling Policy, with a system card noting expanded agentic-risk and thinking-budget evaluations; the visible reasoning was explicitly framed as a transparency experiment whose faithfulness caveats Anthropic itself published.13
Lineage
Predecessor: Claude 3.5 Sonnet (October 2024 version). Successor: Claude 4 (May 2025). See Claude (model family).
