Five Reasoning Levels, One Model: How to Use the Effort Dial

GPT-6 Astra's five selectable reasoning levels are the most consequential new API primitive of the release — one model that can be a fast responder or a deep thinker, per request.

Five Reasoning Levels, One Model: How to Use the Effort Dial

Buried in GPT-6 Astra's launch spec is the feature with the biggest practical consequence: five selectable reasoning levels. Instead of choosing between a fast model and a smart model, developers now set the effort dial per request — and that changes how you architect against the API.

What the levels do. Each step up buys deeper deliberation — more internal computation before an answer — at the cost of latency and tokens. The lowest levels behave like a responsive assistant for lookup-class questions; the highest levels approach the deliberate, multi-step reasoning that frontier benchmarks measure. OpenAI's 96.0% GPQA Diamond result reflects the ceiling setting.

Why it matters architecturally. The old pattern was routing between model tiers — a cheap model for most traffic, a premium model escalated selectively. The effort dial collapses that into one integration: same model identity, same tool surface, same behavioral baseline, with cost adjusted per call. For agentic systems, where a single session interleaves trivial steps and hard ones, per-step effort control is a cleaner primitive than model switching.

How to set it in practice. Start from the bottom and escalate on signals: complexity of the question, stakes of the answer, presence of tool results requiring synthesis. Classification and extraction stay at low levels; multi-document reasoning, planning, and code review get the higher settings. Log the level used alongside outcomes — your own traffic will teach you the mapping faster than any guide.

The caveat. Effort is not a guarantee. Higher levels raise the ceiling but also the variance in latency and cost, and some failure modes — confidently wrong synthesis — can persist at every setting. The dial is a budget instrument, not a correctness guarantee, and evaluation suites should exercise the settings you actually ship.

Expect every frontier lab to ship an equivalent primitive. Tunable reasoning is the rare feature that helps vendors and customers simultaneously — better economics for them, better cost control for everyone else.