What is it?
Reasoning costs tokens. Many reasoning models therefore let you choose how hard they think — for example a low/medium/high setting or a maximum token allowance for thinking. More thinking can help on hard questions but makes answers slower and more expensive.
Also known as: reasoning effort, thinking budget, budget tokens, max reasoning tokens, reasoning effort setting, low medium high reasoning, extended thinking settings.
Technical definition
Providers expose controls such as effort levels or maximum reasoning-token budgets; names, semantics and defaults differ per provider and model and are subject to change. Reasoning tokens usually count toward output limits and billing even when not returned (Reasoning Transparency and Hidden Reasoning). Effective use requires measuring the accuracy-latency-cost frontier on your own tasks (How to Evaluate AI Systems and Outputs).
Entity type: technique · Freshness class: volatile.
When should I use it?
Use whenever you deploy a reasoning model and care about cost or latency.
Implementation patterns
- Default low, escalate on failure: retry with higher effort only if a check fails.
- Per-route settings: low for chat, high for code review.
- Timeouts and streaming for UX.
- Cost guardrails: max tokens per request and per user.
Technologies and frameworks
Commonly associated: provider API reasoning parameters, usage / token accounting fields. Names are examples of what exists, not recommendations, and projects change quickly.
Practical guide
- Find the provider's current parameter names in its docs.
- Benchmark each setting on your evaluation set.
- Pick the cheapest setting meeting your quality bar.
- Set output limits with headroom.
- Monitor cost and latency in production (LLM Observability (Tracing, Logging and Monitoring)).
Common failure modes
- Budget too small, causing weak answers.
- Output limit smaller than reasoning plus answer, causing truncation.
- One global setting for very different tasks.
- Cost surprises because reasoning tokens are not visible.
Troubleshooting
- Truncated or empty answers — Raise max output tokens; confirm how your provider counts reasoning toward it.
- Bills higher than expected — Read usage fields for reasoning-token counts; lower effort.
- Accuracy dips after lowering effort — Raise it only for the task types that regressed.
Contrasted with
- Test-Time Compute (Inference-Time Scaling) — Test-time compute is the general idea; a thinking budget is one provider control for it.
Questions this page answers
- What is a thinking budget?
- How do I set reasoning effort?
- Why is my reasoning model answer cut off?
- How do I reduce reasoning model cost?
Typical goals: control reasoning cost; tune latency of reasoning models; avoid truncated answers on reasoning models.
Sources and verification
Status: not independently verified. This entry was authored on 2026-10-06 from general technical knowledge, without live checking of the sources below. Treat every source as a lead to confirm (identifier, title, authorship, current version) before publication or citation.
Freshness class: volatile — Products, versions, protocols, benchmarks, leaderboards, regulations and dates. Verify against the primary source before relying on specifics; re-review every 3 months.
Time-sensitive claims to verify before relying on them:
- Parameter names, allowed values, defaults and billing treatment, which differ by provider and change.
Canonical sources (unverified leads):
- Provider API documentation for reasoning / thinking parameters (identify current page for the model you use) — locator to be identified