OpenAI’s Decisions API is described as a limited-preview service for asking Luna to choose among predefined answers, while a new Hugging Face community guide offers developers a practical way to think through building decision workflows. The key distinction: the guide says the announcement does not establish a public endpoint, request format, price or general availability.
OpenAI Watch analysis
What happened
The guide says OpenAI introduced Decisions API in its 29 September DevDay recap, describing Luna as answering user-defined questions from a finite set of choices, using text or image context. It suggests uses such as classifying support requests, routing work and choosing an agent’s next step.
For developers, the article lays out a useful implementation approach: define the decision and allowed answers first, assemble representative evaluation examples, make ambiguous cases visible, and set explicit approval boundaries. It also offers an example using the separate DecisionApi service and its Jev model, while warning that this is not a Luna demonstration and that a DecisionApi subscription does not grant OpenAI access. Read the Hugging Face guide.
Why it matters
A fixed set of answers can make a model easier to connect to software than free-form prose, because downstream code can map a label to a known action. But a neatly constrained answer can still be wrong. The guide’s practical advice is to decide what happens when evidence is missing, test against realistic and ambiguous cases, and keep consequential actions under explicit control.
It also helps avoid a familiar launch-day muddle: an announced product, an account with access, and a workflow tested in production are three different things. The guide says the public announcement does not provide enough detail to treat Luna as an API developers can simply call today.
Our read
This is a useful engineering guide precisely because it separates the announced OpenAI service from a third-party example. Start with the decision contract and the evaluation set, not a guessed endpoint or a model name that happens to look close enough. The plumbing can wait until authorised documentation tells you what actually works.
What to watch
- Whether OpenAI publishes Luna-specific API documentation, including a callable model identifier, request format and limits.
- Whether access expands beyond the limited preview, and what pricing and latency terms apply.
- How developers handle ambiguous cases and verify that a model’s allowed answer is the right one.
Discussion spark: For a workflow that can trigger real actions, is a model limited to predefined answers safer in practice, or does that simply make a wrong decision easier to automate?
Sources and evidence
- OpenAI Decisions API Luna Model: A Practical Guide (5 October 2026, 02:27 UTC)
OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.