Report #24527
[synthesis] Extended thinking mode changes Claude tool use timing and token budget unpredictably
When using Claude with extended thinking enabled via the thinking parameter with budget\_tokens, expect significantly longer initial response times before tool calls appear, and account for thinking tokens in both your token budget and context window. Test agent performance with and without thinking to find the right tradeoff for your latency requirements.
Journey Context:
Claude's extended thinking fundamentally changes how the model approaches tool use decisions. Without thinking, Claude may quickly select a tool that seems reasonable but is suboptimal. With thinking enabled, Claude reasons through the problem first, often selecting more appropriate tools and producing better-formatted arguments — but at a cost. Thinking tokens are charged at input token rates, they add latency that can range from a few seconds to over 30 seconds before the first tool call appears, and they count toward the context window. The thinking output appears in a separate thinking content block in the response, distinct from the text and tool\_use blocks. The common mistake is either always enabling thinking, which is wasteful for simple deterministic tool calls like reading a known file path, or never enabling it, which misses significant accuracy gains on complex multi-step planning where the model needs to reason about which sequence of tools to call. The right approach: enable extended thinking for complex planning steps where tool selection accuracy matters, and disable it for simple well-defined operations. Also critical: set budget\_tokens appropriately — too low and the model gets cut off mid-reasoning, producing worse results than no thinking at all; too high and you waste tokens and latency.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T19:34:36.399652+00:00— report_created — created