OpenAI made the Decisions API generally available on October 22, 2026. The company took the endpoint, which runs on GPT-6 Luna, out of public beta. The API gives applications constrained, typed responses instead of free-form text: a probability, one of the predefined choices, or a rubric score.
Pricing is $0.10 per million input tokens. For cached input, the price drops to $0.01; output tokens are free. OpenAI says the endpoint returns results up to 10x faster than asking the same questions to GPT-6 Luna through the Responses API.
Three question types, one endpoint
Requests are sent to POST /v1/decisions, and the only currently supported model is gpt-6-luna. Each request contains 1 to 400 questions along with a shared input (text or message, and tool-call and tool-output arrays). The response comes back as an answers array that preserves the names the developer gave to the questions.
There are three question types. The Predicate type returns the probability that a condition is true, between 0 and 1; for example, whether there is visible damage in a product photo. The Choice type picks one of the choices defined by the developer and returns a confidence value along with a probability for each choice. The Score type scores the input against ordered rubric levels from 2 to 10 and returns the probability-weighted average of the level indices, so the score can land between levels.
Structured data can be sent directly with the input_json part in the user message. Images can be provided as base64 data URLs or from a public HTTP or HTTPS URL. Files, audio, and item references are not supported, and file_id input is not accepted. If a decision needs to follow a response, the docs recommend sending separate requests; if you need a custom JSON object or the model to call tools, it points you to Structured Outputs or function calling in the Responses API.
What was added after the beta
After the beta period, OpenAI added GPT-Live integration. Voice apps can run decisions as the transcript arrives, and Live emits a decision event for the app to handle. The company’s sample setup picks an avatar’s facial expression from transcript updates.
Prompt caching arrived for reused context. A shared input prefix at least 1,024 tokens long can be read from cache in repeat requests; question content and question-specific images stay outside that prefix. There is also inference residency in the US and Europe, extra capacity in several regions in Asia, local JSON input, image attachments to question instructions, and a limit that has risen to 400 questions per request. The previous limit was 200 questions. Hosted HTTP and HTTPS image inputs, conversation-history evaluation with tool calls, organization usage reporting, and Private Link support are also on the list.
Pricing and data controls
Regional processing premiums and long-context input multipliers apply. Cache read and write fees are the same as the standard gpt-6-luna prompt caching rates. Regional processing is supported in the US and Europe (EEA and Switzerland), and data residency is available in all supported API regions. Zero Data Retention and HIPAA usage are available for eligible customers. The usage endpoint returns token usage and request counts by project, API key, or model when called with an admin key that has the api.usage.read permission. Developers can try out decisions without writing code in the Decisions Playground.
What came out of customer tests
OpenAI says its own teams use the API in products and workflows: the personal finance team classifies transaction descriptions, and the communications team sorts news by topic and scores whether an article supports the main narrative.
Customers quoted in the announcement shared the gains from their own evaluations. Box said it sped up routing decisions by about 5x in 30 insurance certificate cases while maintaining the same accuracy as the Responses API, and that it sends low-confidence cases to human review. Datadog said it scores the risk of rshell commands generated by agents before running them and gets results 40x faster at p95 latency than frontier models. Ramp said it does AI cost-to-job attribution 4x faster and 30% cheaper compared with GPT-6 Luna. Sierra said it selects the next step in live customer conversations with p95 latency under 280 milliseconds, reporting that cost per call dropped by 90%. Speak said that with the GPT-Live integration, it cut time to next action from 1.4 seconds to 0.13 seconds.
Why it matters
In agent workflows, not every step needs to produce a long language model response. Decisions like classification, routing, and prioritization can be resolved with a single typed response. OpenAI’s move aims to reduce both latency and token cost by moving these intermediate decisions to a separate, cheaper endpoint.
Source: Unite.AI

Share your experience
Bir sorunun, farklı bir çözümün veya eklemek istediğin bir notun var mı? Giriş yapmadan da yazabilirsin.