Source: Claude's Extended Thinking (Anthropic, 2025)
Anthropic's launch post for extended thinking in Claude 3.7 Sonnet: serial vs parallel test-time compute and why the thought process is shown.
Contents
- What: research/news post
- Publisher: Anthropic
- Date: 2025-02-24
- Original: Claude’s extended thinking
Summary
Introduces extended thinking with Claude 3.7 Sonnet as the same model given more time, not a separate model. It separates serial test-time compute (longer reasoning) from parallel test-time compute (many samples plus selection), and discusses the benefits and risks of showing the reasoning.
Key claims
- Extended thinking is a mode of the same model; developers set a thinking budget.
- Serial scaling: AIME accuracy rises roughly logarithmically with allowed thinking tokens.
- Parallel scaling: with the equivalent of 256 samples, a learned scorer and a 64K budget, Claude 3.7 Sonnet reached 84.8% on GPQA. This was a research result, not the deployed product.
- Showing the reasoning helps trust and alignment research, but it may be unfaithful to what drives the answer, it is not character-trained, and it may help jailbreakers.
- Gains also show up in agentic settings (OSWorld, Pokémon Red), where the advantage grows with more steps.