OpenAI Decisions API Is GA: 400 Questions a Call, Cached Input

The OpenAI Decisions API is now generally available. It returns a probability, a pick from a fixed list, or a score, instead of free text. OpenAI announced it at DevDay as a limited preview, then opened a public beta on October 6. The GA release is more than a label change. You can now ask 400 questions in one call, and the shared context can be cached. For anyone running classification inside an agent, that changes the bill.

What actually happened

Per the OpenAI API changelog entry dated October 22 and the Decisions guide:

• Model: gpt-6-luna on the v1/decisions endpoint.
• Question cap: up to 400 questions per request, double the 200 allowed in beta.
• Three answer types: predicate (probability a statement is true), choice (one of your options, with probabilities and confidence) and score (a rating against 2 to 10 levels).
• New since beta: prompt caching on shared input, images inside individual question instructions, structured data through input_json, hosted image URLs, conversation history with tool calls, org usage reporting and Private Link.
• Pricing: $0.10 per million uncached input tokens. Output is not billed. Cache hits get Luna's standard cached rate, and cache writes cost extra. Regional and long-context premiums can apply.
• Speed: OpenAI says about 10x faster than asking Luna the same thing through the Responses API. It does not publish the test conditions.

Why the Decisions API matters for your business

Batching is the real price cut. Picture one support ticket you need to check for urgency, refund intent, language, product line and churn risk. That used to be five calls. Now it is one call with five questions, and the ticket text is billed once. With caching, a system prompt you reuse all day gets cheaper on every hit.

JSON in means less glue code. Send the order record as structured data instead of stuffing it into a prompt string. Fewer templates, fewer bugs.

Confidence scores are your human off-ramp. Every choice and score comes back with confidence. Pick a threshold, and anything below it goes to a person. That is the difference between automation and an expensive guess.

Still benchmark before you switch. Early beta testers on OpenAI's forum reported mixed accuracy against TypeSafe's Jev. Run 200 of your own labeled records through it and compare against what you use now.

Sources: OpenAI API changelog, OpenAI Decisions guide, OpenAI Developer Community.