Chain-of-thought and extended thinking

Getting the model to reason before answering, and how that is configured now.

4 min read · Lesson 12 of 12 in this domain

Asking a model to reason step by step before committing to an answer improves accuracy on multi-step problems, and there are two distinct ways to get it. Prompted reasoning is just instruction — tell the model to work through the problem before answering — and it remains useful and cheap. Extended thinking is a model capability where reasoning happens in dedicated thinking blocks. The configuration for that has changed: fixed budget_tokens is removed on current models, replaced by adaptive thinking, where Claude decides how much to think and you tune depth with effort. The raw chain of thought is never returned; you can opt into a summary.

• Adaptive thinking (thinking: {type: "adaptive"}) replaces fixed budget_tokens, which returns a 400 on current models.
• Depth and overall spend are tuned with effort inside output_config, not with a token budget.
• Thinking interleaves with tool calls, so the model can reason between steps rather than only up front.
• The raw chain of thought is never returned. A summarised view is available via the display option; the default omits it.
• Pass thinking blocks back unchanged when continuing on the same model; editing or reconstructing them breaks the turn.
• Prompted step-by-step reasoning still works and costs nothing extra — it is not mutually exclusive with the capability.

Source: Claude Docs — Adaptive thinking · Independent study aid, not affiliated with or endorsed by Anthropic.