Reasoning and effort controls
Reasoning models expose knobs (thinking budgets, effort levels) that trade latency and cost for accuracy.
Changes fast: check the latest updates too. Reviewed 17 Oct 2026.
Key ideas
- Use low effort for simple lookups and formatting, high effort for planning, math, code and multi-step tool use.
- Interleaved thinking lets a model reason between tool calls in an agent loop.
- Thinking tokens are billed as output; set budgets per route, not globally.
- Do not over-specify step-by-step instructions to reasoning models; describe the goal and constraints.
Common mistakes
- Running every request at maximum effort; latency and cost multiply with little gain on easy tasks.