How Do You See Claude's Thinking, and Can You Trust It?
In the Claude apps, expand the Thinking section above a response. What you see is a summary of Claude's approach, not its raw reasoning. Use it to catch a wrong assumption early, but don't treat it as proof the answer is right.
Answer this: [your question].
Then add a short explanation with:
- the 3 facts or assumptions your answer depends on
- the evidence for each, quoting [the document or data I gave you]
- what would have to be different for your answer to change
For API calls, turn on the summary with this setting:
"thinking": { "type": "adaptive", "display": "summarized" }
- 1Full thinkingbilled, never shown
- 2Summary you can expandcheck assumptions here
- 3The answerstill needs its own check
Try it now
- 1Ask Claude something with a real trade-off, such as a decision you're weighing this week.
- 2While it works, watch the Thinking indicator and its timer. When it finishes, expand the Thinking section above the response.
- 3Find one assumption in the summary that you know is wrong or incomplete, and correct it in your next message.
- 4You'll know it worked if the new answer changes where you expected it to. If nothing changes, the conclusion didn't depend on that assumption.
Use it when
Use it when an answer seems off and you want to know where it turned, on high-stakes work before you act on it, and when you're tuning a prompt and want to see which instruction Claude latched onto.
Skip it when
Skip it on routine work where you can judge the result directly. Reading a summary of a good rewrite tells you nothing. And don't use it as an audit trail for compliance, since it's a summary and not a record of how the answer was reached.
Why it works
Anthropic's docs say what you see is never the raw chain of thought. The text in a thinking block is a summary. On the API, display defaults to "omitted" on current models, including Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1, so the thinking field comes back empty until you opt in with thinking: {"type": "adaptive", "display": "summarized"}. Either way you're billed for the full thinking Claude generated, not the summary you see, so the visible text is shorter than what you paid for.
Example
You ask whether to raise prices 10% for existing customers on a date you give, and paste your churn numbers. The summary shows Claude treated all customers as being on monthly plans, but 60% of yours are on annual contracts. You reply with the contract split, and it can redo the answer. You found that in the summary before the wrong number reached your deck.
Going further
Don't try to pull the reasoning into the answer. On Opus 5.5, Sonnet 5.5, Fable 5.1 and some others, Anthropic says a prompt asking for internal reasoning in the output may be refused with stop_details.category: "reasoning_extraction". Triggers include a <thinking> or scratchpad section, a reasoning or trace field in JSON or a tool input, and asking for reasoning verbatim. The wording can sit in a system prompt, skill or tool description. Anthropic says this refusal can be billed even before any output, and has no recommended fallback model, so change the prompt. To measure cost, read usage.output_tokens_details.thinking_tokens.
Common mistakes
- Reading the summary as Claude's exact thought process, then quoting it as evidence.
- Adding a
reasoningfield to a JSON schema and getting refusals. Ask for a shortexplanationor the supporting evidence instead. - Assuming an empty
thinkingfield on the API means Claude didn't think. On current models the default display hides it, and it's still billed. - Building code that expects every response to start with a thinking block. Anthropic's docs say a turn where Claude skips thinking has none.
Sources
- Thinking, Claude Platform Docs
- Steering thinking, Claude Platform Docs
- Refusals and fallback, Claude Platform Docs
- Thinking and effort in Claude, Claude Help Center
- Define tools, Claude Platform Docs
Checked 6 November 2026. Products change often, so confirm details with the vendor.