OpenAI opened the Decisions beta on October 6. For text or images it returns a condition estimate, a choice or a score. Current documentation lists GPT-6 Luna; roughly tenfold speed compared with Responses is a vendor claim, not our measurement.

Possible use: customer support could route requests first and leave ambiguous cases to people. A service company might test preliminary sorting of damage photographs. The useful change is not permission for AI to decide everything, but a small step with predefined options that can be tested separately.

Requirements include real examples with verified outcomes, measurements of missed defects and false alarms, and a way to decline a decision. A probability-like number is not automatically well calibrated. When an incorrect choice triggers payment, changes a machine or rejects an important application, safeguards must exist outside the model.

Optimistic editorial scenario: a team with an existing API integration might prepare a limited pilot in 2–6 weeks. Initially it would compare decisions against human work without taking automatic action. Beta status means interfaces and availability may change; production timing depends on local testing, not a successful demonstration.