OpenAI's Decisions API Is Now GA. Use It for Bounded Answers, Not Every AI Task

OpenAI moved the Decisions API to general availability

OpenAI's October 22 announcement says the dedicated Decisions API is generally available with GPT-6 Luna. The endpoint evaluates text, images or structured JSON and returns bounded, typed answers.

The release adds prompt caching, native JSON input, images inside individual question instructions, organization usage reporting and up to 400 questions in one request. The public beta allowed 200.

This is a production availability change, not proof that every classification should move to a model. OpenAI reports that Decisions can answer up to 10 times faster than GPT-6 Luna through the Responses API, but teams must test that claim on their own inputs and regions.

The practical opportunity is narrower: products that currently generate prose only to parse a category, boolean or score can replace that detour with a response shape designed for decisions.

Choose predicate, choice or score from the product question

The API supports three question types. A predicate estimates whether a condition is true. A choice selects one supplied option. A score places evidence against ordered levels and can return a value between them.

For a support ticket, a predicate can estimate whether the customer reports visible damage. A choice can route the case to delivery, billing or product support. A score can rate urgency against clearly described levels.

OpenAI's Decisions guide recommends observable criteria, distinct choices and levels whose adjacent meanings are clear. “Is this important?” is weak. “Does the ticket report a safety risk requiring review today?” can be labelled and tested.

If the product needs extracted fields or a written explanation, use Structured Outputs through the Responses API. If it needs the model to request a tool with arguments, use function calling. A typed answer is not a tiny general-purpose agent.

Probabilities need business thresholds before they trigger action

The API returns probability information, but the application decides what happens next. That boundary is the product decision.

Imagine a damage predicate returning 0.72. An automatic replacement may be reasonable for a low-value item with reversible fulfilment. The same probability should not approve a large payout or reject a safety report without review.

Build a labelled set from real examples and separate false-positive from false-negative costs. The guide tells teams to choose thresholds from those costs rather than treating one confidence value as universally safe.

Record the input version, question, answer, threshold, route and human override. An AI product audit should later be able to reconstruct why a case was automated, reviewed or refused.

Batch only independent questions about shared evidence

The GA endpoint accepts between one and 400 questions about the same input. Batching is useful when each answer can be evaluated independently against shared evidence.

A product photo can be checked for visible damage, category and packaging condition in one request. If the second question depends on the first answer, OpenAI says to use a separate request.

That distinction prevents a large question array from becoming hidden workflow logic. Keep dependencies in application code where the team can test, version and observe them.

Prompt caching can reuse a matching shared-input prefix when it contains at least 1,024 tokens. Cache economics matter for repeated reference material, but do not pad evidence just to cross a threshold. Smaller, relevant input is easier to reason about.

The GA release still has visible limits

GPT-6 Luna is the only model currently supported by the Decisions API. The API reference also limits questions to predicate, choice and score shapes, with a refusal possible for any individual question.

Conversation history and tool outputs can be supplied as evidence, but the API does not execute an action. Files, audio and item references are not supported in the current decision input described by the guide.

OpenAI lists a current price of $0.10 per million uncached input tokens and no output-token charge for `/v1/decisions`. Cached input is discounted and cache writes are charged. Pricing and eligibility can change, so product budgets should read the current documentation rather than bake a launch post into code.

General availability removes a beta label. It does not remove probabilistic errors, model dependence or the need for human review when consequences are hard to reverse.

What product teams should test now

• Pick one frequent decision with a small, fixed answer space.
• Write the evidence and observable question in an AI product requirements document.
• Collect labelled normal cases, edge cases and deliberate ambiguities.
• Measure false positives and false negatives separately.
• Set an automation threshold and a human-review band.
• Log the API answer, product rule, downstream action and override.
• Run an AI proof of concept against one risk before changing the production route.

The best first test is not a dramatic autonomous workflow. It is a decision your team already makes often, with examples and a clear consequence when it is wrong.

If users see or correct the result, design that state explicitly. The principles in AI UX design still apply: show uncertainty, preserve control and make recovery obvious.