Extended thinking accuracy improvement
🧠
~50%
Standard
Without thinking
🚀
~90%+
Extended
With thinking
Step 1: Set budget_tokens
Define the maximum thinking budget (minimum 1,024 tokens)
Step 2: Claude reasons internally
Plans, considers alternatives, checks for errors in a scratchpad
Step 3: Calibrates depth
Adaptive thinking (Opus/Sonnet 4.6) adjusts reasoning to problem complexity
Step 4: Produces final answer
A polished response incorporating insights from the reasoning process
Computer use in practice
Computer use works through a screenshot and action loop. Claude takes a screenshot, decides what to click or type, your environment executes the action, and you send back the new screenshot. This loop continues until the task is complete. It works with any application that has a visual interface, including legacy software, web applications, and desktop programs. Currently billed at standard token rates plus vision costs for the screenshots.
🧮Extended thinking: budget_tokens controls reasoning depth
🎯Adaptive thinking: auto calibrates per request (Opus/Sonnet 4.6)
🖥️Computer use: full GUI control via screenshot and action loop
⚡Fast mode: 6x pricing for faster Opus 4.6 output