Claude Extended Thinking Budget: How Many Tokens to Set

Developer documentation for Claude's Extended Thinking notes that thinking tokens have a minimum budget of 1024 tokens, and recommends starting with the minimum budget and increasing incrementally. For workloads where the optimal thinking budget is above 32K tokens, use batch processing to avoid system timeouts and open connection limits. Additionally, prefilling extended thinking is not allowed, and manually changing model output text following its thinking block degrades results.

API Config: thinking={"type": "enabled", "budget_tokens": 16000} Prompt: Evaluate the solvency stress-test scenarios in <portfolio_data>{{DATA}}</portfolio_data>. Consider multiple approaches and show your complete reasoning. Try different methods if your first approach doesn't work.