OpenAI announced general availability of its Decisions API on October 22, 2026. The service uses GPT-6 Luna for questions with bounded answers. Its documented answer types are predicate, choice, and score, allowing applications to estimate a condition, choose an allowed option, or apply an ordered rubric.
Why a smaller decision surface helps
Many workflows do not need a paragraph. They need a routing label, an escalation decision, or a structured quality score. Restricting the result makes downstream validation easier and reduces the temptation to parse business logic out of generated prose.
However, a bounded answer can still be wrong. A valid category is not necessarily the correct category. Probability estimates should be tested for calibration rather than treated as guarantees.
A sensible first use
Consider routing incoming support requests to an approved set of queues. Include an ambiguous or manual-review path. Test real examples that contain missing information, multiple issues, and unfamiliar terminology before allowing the classifier to affect customer response times.
The announcement reports supplier and customer performance claims. Those results are specific to their workloads and should not be presented as a guaranteed speedup for yours. Measure end-to-end latency including retrieval, network time, retries, and any required review.
Keep authority outside the prediction
Use ordinary application rules to decide whether a predicted route is allowed for the current account. Record the selected choice and the policy version that governed the next step. A fast decision service is most useful when the surrounding workflow can verify its input, constrain its authority, and recover from a mistaken answer.
Sources & further reading
Reviewed October 26, 2026. Product availability and regulatory guidance can change. Vendor performance claims are not independent Rovio AI benchmarks.
Found this useful? Share it.