On Tuesday, October 6, 2026, OpenAI opened the Decisions API in public beta, a new endpoint that returns typed answers to developer-defined questions about text and images. According to the API changelog as reported by Unite.AI, the service runs exclusively on the gpt-6-luna model and is offered through a dedicated route, POST /v1/decisions. The company expects it to reach general availability in the coming weeks.
The API is not entirely new. OpenAI introduced it at DevDay 2026, where it was available in a limited preview. At the time, the company explained that the tool applies the Luna model's reasoning to a set of user-defined questions, each with a finite set of possible answers. GPT-6 Luna is a reasoning model that OpenAI launched alongside GPT-6 Sol on September 22, 2026.
The main promise is speed. According to the documentation, the Decisions API produces its answers about 10 times faster than the Responses API, OpenAI's general interface for generating text. The reason is that the model does not have to write a free-form reply: it only needs to estimate probabilities across known options. That makes it suitable for classifying content, routing requests or deciding an agent's next step.
Each request has three parts. The model field selects the evaluating model, which for now can only be gpt-6-luna. The input field carries the shared evidence, either as text or as user messages containing text and images. And the questions array lists what must be evaluated: each question's type, its instructions and the allowed options or levels. Each question carries a unique name, which the API returns alongside its answer.
There are three question types. Predicate questions check a condition, such as whether a product photo shows a crack, and return a probability between 0 and 1; in the guide's example, the illustrative answer is 0.92. Choice questions select a value from fixed options and return the chosen value, the probabilities for each option and a confidence field; the example routes a customer complaint about a double charge to the billing category.
The third type is score questions, which rate an input against ordered levels from lowest to highest, such as the severity of a problem. The result is the probability-weighted average of the level indices, which start at 0, so it can fall between two levels. In the documentation's example, probabilities of 0.1, 0.7 and 0.2 across three levels produce a score of 1.1 with a confidence of 0.55. Answers can also come back as refusals.
The guide includes practical recommendations. OpenAI advises adding a fallback option such as “other” for inputs the categories do not anticipate, keeping distinct concerns in separate questions and defining levels that differ clearly. It also asks developers to calibrate thresholds against labeled examples from their own application, weighing the cost of false positives against that of false negatives. For outputs with a custom JSON schema, it points to Structured Outputs.
There are technical limits. Images must be sent as inline base64 data URLs; hosted HTTP or HTTPS image URLs and file IDs are not supported. Independent questions can share a single request, even with different types, but when one decision depends on another, the application must make separate requests, for example confirming damage first and then asking for the repair category.
Pricing is one of the most striking aspects. Using the endpoint with gpt-6-luna costs $0.10 per million input tokens, and only those tokens are billed: cache reads, cache writes and output tokens carry no charge. Regional processing surcharges and long-context multipliers still apply. Requests to gpt-6-luna outside /v1/decisions keep the model's normal pricing.
On data, the API supports zero data retention and HIPAA-compliant use for eligible customers, with data residency and regional processing in the United States and in Europe, defined as the European Economic Area plus Switzerland. For voice applications, client delegation with the Live API lets an agent choose actions from spoken requests and then report the result back to the user. The API can also be tried in the Playground.
The launch comes amid a surge of interest in so-called decision models. According to TechCrunch, interest took off after TypeSafe AI released Jev in September, followed by decision models from OpenAI and Amazon. Instead of producing text, these models return probabilities over predetermined outcomes, which lets them run faster and cheaper than a conventional language model while keeping the flexibility of the transformer architecture.
On the same Tuesday, the company Musubi unveiled PolicyLM-1.7B, a lightweight open-weight decision model for real-time content moderation. According to TechCrunch, it takes a policy written in plain English and applies it to messages in under 50 milliseconds, at a cost and speed similar to the classifiers social networks already use. Its advantage is that it does not need retraining when the policy changes, letting human teams iterate freely.
“Product teams just want a better understanding of what's happening on their platform, especially as the amount of content is exponentially increasing,” Filip Jankovic, Musubi's co-founder and chief AI officer, told TechCrunch. Jankovic traces his interest back to GLiNER, a 2024 named-entity recognition project that used similar techniques. One early use of these models, the outlet notes, is reining in misbehavior by AI agents themselves.
In conclusion, the Decisions API turns a niche idea into a general-purpose product: closed questions, probability-based answers, low latency and pricing that only charges for input. For developers, the key will be calibrating thresholds carefully and not mistaking a high probability for certainty. For the market, the timing alongside PolicyLM-1.7B shows that decision models now compete both in the cloud and in open versions that anyone can run.
