OpenAI’s Decisions API turns AI outputs into scored choices

OpenAI has released a Decisions API in public beta for applications that need a bounded answer rather than another paragraph. Powered by GPT-6 Luna, it accepts text and images and supports three output forms: a probability that a statement is true, a choice from a predefined set with confidence scores, or a numeric score against a rubric. OpenAI says it can return these typed answers about ten times faster than using Luna through the Responses API.[1][2]

That design addresses a common mismatch. A customer-support workflow may need to decide which queue receives a case. A moderation system may need a probability that content violates one policy. A document pipeline may need a score against a fixed completeness rubric. Open-ended prose can contain the answer, but software must then parse language that can vary from one request to the next. A typed response moves the interface closer to the actual decision.

A probability is not automatically a trustworthy probability

If a model assigns 0.8 to one hundred comparable events, calibration asks whether roughly eighty of those events actually occur. A model can rank risky cases well while being badly calibrated: the highest-risk items may indeed be worse than the rest, yet a displayed 80% may correspond to only 55% in practice. That matters whenever software converts the number into approval, rejection or escalation.

Calibration should be measured on the organization’s own distribution and revisited when that distribution changes. A fraud model evaluated on card payments may not transfer to invoice fraud. A document score tested on clean English forms may drift when a new market introduces scans, handwriting or another language. The typed interface makes monitoring easier, but it does not remove dataset shift.

The threshold belongs to the business, not the model

Consider an illustrative fraud-review system. Suppose review costs £5 per transaction and always prevents the full £500 loss if a case is fraudulent. With calibrated probabilities, no capacity limit and no other consequences, the expected-cost rule is to review when the fraud probability exceeds 1%: the expected avoided loss, p × £500, then exceeds £5. If review prevents only half the expected loss, the threshold rises to 2%. Customer friction, residual losses and limited review capacity also change the policy. The model supplies an estimate; the organization defines the action.

The same logic applies to safety and quality checks. A false positive that delays an ordinary marketing image may be cheap. A false negative that releases a harmful industrial instruction may be severe. Reporting one overall accuracy number conceals those asymmetries. A production dashboard should show false-positive and false-negative rates at the actual operating threshold, plus the outcomes of deferred human review.

Fast answers make governance more—not less—important

A tenfold latency claim can make decisions feasible inside interactive products. It can also multiply the number of automated judgments an organization makes each day. At that scale, small calibration errors become operational patterns. The strongest use of the Decisions API will pair the fast typed output with slow, disciplined measurement: representative test sets, cost-aware thresholds, monitoring by segment and a clearly defined fallback when the system is uncertain.

Report an error or suggest a correction