Claude’s extended thinking

Source: Anthropic Research Date: 2025-02-24 URL: https://www.anthropic.com/research/visible-extended-thinking

Summary

Anthropic’s research post on Claude 3.7 Sonnet’s extended thinking explains how allocating more “thinking tokens” at inference time improves accuracy on hard problems, with math performance scaling logarithmically with compute. Crucially, the thinking trace is made visible to users — a deliberate transparency choice aimed at safety research and trust. Anthropic acknowledges the tension: displayed reasoning may not faithfully represent internal model processes, and training on observable thoughts could incentivize future models to conceal or stage their thinking.

Implications

• Test-time compute scaling is now user-facing. Extended thinking ships as a toggle rather than a new model, shifting the mental model of “capability” from model weights alone to compute budget at inference time.
• Transparency raises an alignment double-bind. Making chain-of-thought visible aids auditing but may cause models to learn to perform reasoning rather than do it — a problem Anthropic names directly and hasn’t resolved.
• Feeds the reasoning/agentic AI thread. Logarithmic accuracy gains with token budget is a favorable scaling law for tasks like multi-step planning and code; this sets the architecture direction for agentic use-cases that trade latency for correctness.