Claude 4 adds extended thinking with tools, interleaved thinking and thought summaries
Claude Opus 4 and Sonnet 4 added "extended thinking with tool use (beta)", in which the model alternates between reasoning and calling tools such as search, and parallel tool use; SWE-bench Verified…
- Date
- 22 May 2025
- Who
- Anthropic
- Confidence
- High
- Deep dive
- Reasoning II, from o1 and o3 to DeepSeek-R1 and the labs that replicated them
Tier: Supporting · Significance: 3/5 · Org(s): Anthropic · Confidence: High Claude Opus 4 and Sonnet 4 added "extended thinking with tool use (beta)", in which the model alternates between reasoning and calling tools such as search, and parallel tool use; SWE-bench Verified was reported at 72.5% (Opus 4) and 72.7% (Sonnet 4), with Opus 4 priced at $15 / $75 per million tokens (announcement). Anthropic's API calls the pattern "interleaved thinking," enabled in 2025 by a beta header and automatic in the adaptive mode of later models (docs). It also introduced "thinking summaries," in which a smaller model condenses lengthy thought processes, which Anthropic said was needed only about 5% of the time. This was the start of the move away from fully raw visible thoughts (B08-06). It came 252 days after o1-preview, 36 after o3's tool-using release. Agentic consequences are covered in B10.