Reasoning Effort
Reasoning effort is an inference-time parameter that controls how many internal computation steps or tokens a model allocates to reasoning before producing an answer, allowing users to trade response latency and cost against output quality on complex tasks.
Reasoning effort is a controllable parameter exposed in the APIs of large reasoning models—those designed to perform extended internal deliberation before responding. It governs the scale of the model's hidden thinking phase, in which the model generates intermediate reasoning steps (variously called scratchpad tokens, chain-of-thought tokens, or thinking tokens) that are processed internally but typically not shown in the final output. Higher effort allocates more computation to this phase, which generally improves performance on tasks requiring multi-step logic, mathematics, code generation, or planning at the cost of increased latency and token consumption.
OpenAI introduced explicit reasoning effort control with the o1 and o3 model families, exposing a `reasoning_effort` API parameter accepting values such as "low," "medium," and "high." Anthropic's extended thinking feature for Claude models exposes a `budget_tokens` parameter that caps the maximum number of internal reasoning tokens the model may generate. Google's Gemini 2.5 thinking models offer analogous thinking-budget controls. At low effort, models typically respond in under a second and consume minimal extra tokens; at high effort, they may generate thousands of hidden reasoning tokens and take tens of seconds before producing a final answer.
The practical significance of reasoning effort is that it makes computational depth a selectable resource. A simple factual lookup benefits little from additional thinking, while a competition-level mathematics problem or a complex multi-file code audit may show substantial quality gains at higher effort. This enables cost-efficient deployments: routine queries are handled cheaply, while complex queries trigger deeper deliberation. OpenAI reported that o3 at high-effort settings outperformed o1 on AIME mathematics benchmarks and achieved top-tier results on competitive programming evaluations.
As of 2026, reasoning effort is a standard inference configuration dimension alongside temperature and maximum output tokens. Research is active on adaptive effort allocation, in which the model or runtime automatically estimates query difficulty and assigns effort accordingly, rather than requiring explicit user specification. Inference providers have also begun pricing reasoning-heavy tiers based on thinking-token consumption, making effort control directly relevant to operational cost management.