OpenAI has launched the public beta of its Decisions API, a new endpoint designed to evaluate text and images to provide typed answers rather than prose. The service currently supports only the gpt-6-luna model, with general availability expected in the coming weeks.
The API processes requests containing shared evidence and a list of specific questions. Each question includes a name, type, and instructions, with the output returned as an array keyed by those names. OpenAI suggests using this tool for tasks involving probabilities, choices, or scores.
Performance tests indicate the Decisions API is approximately 10 times faster than the standard Responses API. During DevDay, decision processing was measured at 150 milliseconds, compared to 1.6 seconds for standard Luna calls. OpenAI has not yet released specific accuracy or calibration data.
Pricing for gpt-6-luna is set at $0.10 per 1 million tokens for input, with no charges for output tokens or cache operations. Regional premiums and multipliers for contexts exceeding 272,000 tokens still apply. Users are advised to calibrate thresholds using their own labeled examples.
The primary competitor is the TypeSafe Jev model, which launched on September 15, 2026. Jev is priced at $0.042 per 1 million tokens and offers response times between 70 and 500 milliseconds. OpenAI maintains a competitive edge through image input support, compliance options, and public beta access.
