OpenAI's Decisions API enters public beta on gpt-6-luna at $0.10 per 1M input tokens

OpenAI opened its Decisions API to public beta on October 6, returning yes/no probabilities, picks and scores; early tests favor rival Jev.

OpenAI's Decisions API is now in public beta. The dedicated POST /v1/decisions endpoint returns typed answers instead of prose: a probability that a condition is true (predicate), one option from a fixed set (choice), or a rating against ordered levels (score). If you run classification, ticket routing or eval gates, it is a cheap call to test. According to OpenAI's guide, gpt-6-luna is the only supported model. OpenAI expects general availability "in the coming weeks" and has not given a date.

Pricing is $0.10 per 1M input tokens. OpenAI says there are no output, cache-read or cache-write charges, though regional processing premiums and long-context multipliers still apply, Unite.AI reports. OpenAI says the endpoint is "about 10x faster than the Responses API." MIXED notes that OpenAI published no benchmark to support that claim. The guide requires images as inline base64 data URLs. If one question depends on another's answer, you need a second request.

Early independent results are mixed. In Arize's benchmark, OpenAI scored 81.5% balanced accuracy on RAGTruth hallucination detection, against 85.5% for TypeSafe Jev. Arize says that looks like a clear gap, but the confidence intervals touch, so it calls accuracy a draw. Arize cautions that its prompt layout for OpenAI's input format could affect those scores. A developer in the OpenAI community thread found Jev cheaper, faster and more accurate on judgment-heavy questions, based on only a few hundred game questions. The same thread reports that choice questions overweight the likeliest option: a coin biased to land heads 70% of the time came back as heads 98% of the time.

Open-weight alternatives also exist. TokenCost lists Perplexity's pplx-decider at $0.04 per 1M input tokens. Before you switch, calibrate thresholds on your own labeled data.

Sources