Outcome. You can distinguish visible explanations, hidden reasoning tokens, and extra inference effort, then decide when additional reasoning cost is justified.
Models can improve on some multi-step tasks by using additional inference computation before producing a final response. APIs may expose this as a reasoning model, a thinking budget, or an effort level. The implementation and accounting differ by provider, but the product trade-off is durable: more computation can improve difficult math, coding, planning, or constraint satisfaction while increasing latency and cost.
Do not equate a fluent rationale with a faithful transcript of the mechanism that caused the answer. A model can produce a convincing explanation after making a choice for other reasons, and providers may intentionally hide or summarize internal reasoning. Evaluate the final answer, evidence, tool trace, and task-specific checks; do not use chain-of-thought prose as a security proof.
Reasoning effort should be routed by task. Simple extraction, formatting, or classification may not benefit. Hard boundary cases may. Compare effort levels on the same dataset and use the cheapest level that meets the required quality and safety threshold.
Mental model. Reasoning effort buys additional inference work; only an eval can tell you whether that work improved this task.
Evidence trail — reviewed 23 July 2026. Chain-of-thought prompting evidence: https://arxiv.org/abs/2201.11903. Research on unfaithful explanations: https://arxiv.org/abs/2305.04388. Use current provider documentation for how reasoning tokens, budgets, and summaries are exposed.
Build 12 cases: four simple extractions, four moderate constraint problems, and four difficult planning or code-debugging cases. Run supported low/medium/high effort modes. Score correctness, constraint compliance, latency, and total tokens. Set a routing rule before looking at production traffic.