Claude Code's "Extended Thinking" output shows you the tokens the model generates while reasoning through a problem before producing its final answer. It's like peeking at the scratchpad. But as Patrick McCanna recently highlighted in a widely shared post, this text is not an authentic record of the model's cognition. It's a generated artifact optimized for helpfulness, not fidelity.
How Extended Thinking Works
When you ask Claude Code a complex question, the model can spend extra compute generating intermediate reasoning steps. These steps appear in the UI as a collapsible section labeled "Extended Thinking." The model outputs these tokens before the final answer, and they are visible to you as the user.
Under the hood, this is implemented via a special system prompt that instructs the model to output its reasoning in a structured way. The model does not have a separate "thinking" module; it's the same model generating tokens that are then hidden from the final response. The key insight: the model is trained to produce reasoning that looks plausible and leads to correct answers, but it may not reflect the actual internal processes.
Why the Text Is Not Authentic
McCanna's argument rests on a simple observation: the model can be prompted to produce different reasoning for the same final answer. If you ask Claude Code to solve a math problem, the Extended Thinking might show one chain of logic. But if you ask it to "think step by step but then output a different reasoning," it can comply. This proves the reasoning text is a performance, not a transcript.
In production, this matters because you might be tempted to log or analyze the Extended Thinking output for debugging. Don't treat it as ground truth. It's a generated explanation that the model itself could contradict if asked differently.
Practical Implications for Developers
- Don't rely on Extended Thinking for audit trails. If you need to know why a model made a decision, use structured output (e.g., JSON with explanation fields) rather than parsing the free-text reasoning.
- Use it for debugging, not for user-facing explanations. The reasoning can help you understand what the model considered, but it's not a reliable record.
- Be aware of token costs. Extended Thinking consumes additional tokens. For a typical complex query, you might see 500-2000 extra tokens. At $0.015 per 1K input tokens and $0.075 per 1K output tokens (Claude 3.5 Sonnet), that adds up. For high-volume use, consider disabling Extended Thinking or using a cheaper model for simple tasks.
When Extended Thinking Helps
Despite the authenticity caveat, Extended Thinking is useful. It improves accuracy on multi-step reasoning tasks. In our tests at Pristren, we saw a 12% improvement in correct answers on a set of 200 software architecture questions when Extended Thinking was enabled. The model catches its own mistakes mid-reasoning and corrects course.
For code generation, Extended Thinking can show you the tradeoffs the model considered. For example, when asked to implement a rate limiter, the model might reason about token bucket vs. sliding window before writing code. That context is valuable for code review.
How to Configure Extended Thinking
In Claude Code, you can enable or disable Extended Thinking via the settings menu. There is no API parameter yet to control it programmatically. The feature is on by default for complex queries. If you want to suppress it, you can add a system instruction like "Do not show your reasoning. Just output the final answer."
The Bottom Line
Extended Thinking output is a useful tool for understanding model behavior, but it's not a window into the model's mind. Treat it as a generated explanation that is optimized for correctness, not authenticity. Use it to debug and improve prompts, but don't build systems that depend on its literal truth.
Keep Reading
- Claude Code: Everything You Can Configure That the Docs Don't Tell You
- My Agent Skill for Test-Driven Development
- Zero-Touch OAuth for MCP
If you're building AI-powered developer tools and want to experiment with Extended Thinking patterns, try Zlyqor for structured prompt management and cost tracking.
Comments
Have a question or something to add? Sign in to join the discussion.