The Pulse
OpenAI Launches Decisions API, Claiming 10x Faster Typed Answers
OpenAI released the Decisions API in public beta on October 6, powered by GPT-6 Luna. The endpoint returns probabilities, fixed-option choices or rubric scores from text and inline images, and OpenAI says it is up to 10 times faster than th
AI.info Team · Written with AI ·
OpenAI says its new Decisions API can return typed answers up to 10 times faster than its Responses API, but the speed claim comes with a narrower job description: the new endpoint selects among structured outcomes rather than generating open-ended responses. The company released the API in public beta on October 6, powered exclusively by GPT-6 Luna. OpenAI’s announcement says: “Let your app choose the right model, tool, or action in near-real time with Decisions API, now available to all developers in public beta.”
GPT-6 Luna answers three kinds of questions
Developers send shared evidence and an ordered list of questions to POST /v1/decisions. The endpoint accepts text and images, then returns one of three defined answer types: a predicate estimating whether a condition is true, a choice from developer-supplied options, or a score against ordered levels. A request can ask several independent questions about the same input, and the response returns answers in question order.
That structure targets tasks such as routing a support request to a department, checking a product photo for visible damage, or assigning a severity rating to an issue. A predicate returns a probability; choices and scores include probabilities across their available options and a confidence value. For scores, the result is a probability-weighted average of level indices, so it can fall between the named levels rather than matching one exactly.
The speed claim does not make it a general response API
OpenAI’s API guide describes the endpoint as roughly 10 times faster than the Responses API. The changelog and guide do not provide a benchmark methodology, test workload or latency figures behind that comparison. The claim should therefore be read as OpenAI’s product comparison, not as a result independently established across applications.
The tradeoff is a more constrained interface. Only GPT-6 Luna is available, and the endpoint is designed for predicates, fixed choices and ordered scores. OpenAI recommends the Responses API’s Structured Outputs when an application needs an object matching its own JSON schema, and function calling when a model needs to request a tool call with arguments.
Images must be sent inline
Requests can include user messages containing text and images, but the API reference rules out files, audio, function calls and item references. Images must be supplied as base64 data URLs; hosted image links and file IDs are not accepted. That requirement may affect applications that already store images remotely and would otherwise pass a reference rather than encode the image in each request.
The guide also advises developers to set routing thresholds using labeled examples and to account for the cost of false positives and false negatives. OpenAI says the returned probabilities and confidence values should inform application decisions, not replace a threshold chosen and tested for the particular workflow. Its examples include a fallback category for inputs that do not fit the choices provided.
Input tokens set the Decisions API bill
For Decisions requests using GPT-6 Luna, OpenAI lists a price of $0.10 per million input tokens and says there are no charges for cache reads, cache writes or output tokens. Regional-processing premiums and long-context pricing multipliers still apply. The API guide says the endpoint supports Zero Data Retention and HIPAA use for eligible customers, with data residency and regional processing available in the United States and Europe; OpenAI’s data-controls documentation sets out the applicable limits.
The beta is open to all developers, while general availability is still ahead: OpenAI’s guide says it expects to reach that stage in the coming weeks. For now, teams choosing the endpoint must work within its one-model, three-answer-type interface and send images inline.