OpenAI Decisions API vs Jev: what is confirmed
Compare OpenAI Decisions API and TypeSafe Jev using confirmed launch details, Jev's published API, and the open questions about pricing, confidence and images.
OpenAI Decisions API and TypeSafe Jev address the same kind of problem: make a bounded judgment that software can use. OpenAI announced Decisions API at DevDay 2026 as a limited preview that focuses Luna on user-defined questions with finite, predefined answers. Developers can provide text or images and use the answers to classify content, route requests or choose an agent's next action. Jev is already documented as an API for typed questions over supplied state, with probability distributions that code can inspect. That is enough to compare their announced interfaces. It is not enough to declare a winner on accuracy, latency or price.
This comparison is dated September 29, 2026. OpenAI says a broad release is planned in the coming days; a published Decisions API reference could change what is knowable here.
Decisions API vs Jev at launch
| Question | OpenAI Decisions API | TypeSafe Jev |
|---|---|---|
| Public status | Limited preview, with a broad release planned in the coming days | Public early access with published documentation and an API reference |
| Model relationship | OpenAI says the API focuses Luna's intelligence on bounded questions | TypeSafe describes Jev as a purpose-built System One model |
| Input context | Text or images, explicitly named in OpenAI's announcement | state is a string, object or array in TypeSafe's published request schema; native image input is not documented there |
| Answer space | Finite answers defined in advance | Choice, Score and Noul questions with defined answer shapes |
| Probabilities and confidence | The DevDay recap does not specify response fields | Choice and Score include probabilities and confidence; Noul returns a probability of yes |
| Multiple questions | OpenAI names a set of user-defined questions; batching behavior is not specified | Questions can be mixed in one request and are evaluated independently against the same state |
| Price | No Decisions API price in the DevDay recap | TypeSafe's September 15 launch post lists $0.042 per million input tokens and no output-token charge |
OpenAI's DevDay recap, TypeSafe's introduction, API schema and Jev launch post are the sources for this table. “Not specified” means the source does not settle the point. It is not a claim that the feature is absent.
What did OpenAI actually announce?
OpenAI's description is short and specific. Developers define questions and finite possible answers, supply context as text or images, and receive answers for real-time classification, request routing or selecting an agent's next action. It names Luna as the intelligence focused by the API. OpenAI says limited preview is available on launch day and a wider release is planned soon. That supports a clear use case: replace a free-form response with a bounded decision your application can branch on.
The recap does not publish a Decisions API endpoint, request body, response schema, pricing, rate limits or evaluation results. It also does not say whether the response includes a probability for each answer, a separate confidence score, an abstain option, or a way to submit several independent questions in one call. Those details matter when code must decide whether to act, ask a person or send the item to a stronger model. A Luna model price on OpenAI's general API pricing page would not establish the price of this separate Decisions API.
OpenAI explicitly mentions images. That could matter for document routing, screenshots or visual quality checks. The announcement does not yet explain image size limits, preprocessing, or whether the same answer contract applies to text and images. A team building on this should wait for the API reference before designing around those details.
What does Jev publish today?
TypeSafe's Jev documentation describes a different, fully specified contract. The caller sends a state and one or more typed questions. The answers are structured values rather than prose. TypeSafe defines three question types:
| Jev question | What the caller defines | What the API returns |
|---|---|---|
| Choice | Named alternatives and descriptions | A selected option, probability per option and confidence |
| Score | An ordered rubric | A probability-weighted score, probability per level and confidence |
| Noul | A yes-or-no statement | The probability of yes, from 0 to 1 |
The published OpenAPI schema shows state as a string, object or array and the response fields for those question types. TypeSafe says several questions can share the same state in one request while being evaluated independently. Jev does not write a reply or take an action itself; the calling code decides what to do with the values. Our Jev guide covers the primitives, their limits and the published price in more detail.
There is a distinction between a well-formed answer and a correct decision. Jev's Choice result must name one of the supplied options, but the chosen option can still be wrong. Its probabilities and confidence make uncertainty visible; they do not remove the need to evaluate the model on your own data. A confidence threshold should depend on the cost of a wrong action, with an explicit route for uncertain cases.
TypeSafe's launch price is a useful planning number, not a comparable Decisions API quote. OpenAI has not published a Decisions API rate in its announcement. Nor do the two vendors publish a shared test set here. Comparing Jev's token price with the list price of a general Luna call would mix two different products and omit the rest of the workflow.
A fair test: route the same support inbox
Consider 1,000 incoming support messages. The first judgment is which queue should handle each one: billing, product bug, account access, sales or other. A second judgment might ask how urgently a person needs to review it. The application can use those answers to assign a queue, leave uncertain messages for a human, and send the few that need a written reply to a language model or worker. This is the judging half of a workflow, separate from composing and sending the response.
Both products' announced shapes fit the queue choice because the possible answers can be written down in advance. Jev's published interface also supports the urgency rubric and returns distributions for each judgment. OpenAI's recap says its API supports user-defined questions with finite answers, but its precise score and uncertainty fields have not been published. A design that requires a Jev-style confidence value should not assume OpenAI will expose one.
To compare them once Decisions API documentation is available, hold the workflow constant. Use the same message set, category definitions, no-match policy and human-reviewed labels. Then measure the outcomes that affect operations:
| Measure | Why it matters | What to record |
|---|---|---|
| Correct routing | A valid label can still send a customer to the wrong team | Per-category errors, including the “other” bucket |
| Uncertain cases | Automation needs a safe route for weak evidence | Review rate and errors above each proposed action threshold |
| Probability quality | A confidence number helps only if it predicts error on this task | Calibration against outcomes, if both APIs expose comparable scores |
| End-to-end time | A fast model can still sit behind slow retrieval or writes | Time from new message to usable decision and completed action |
| Full cost | Model calls are one part of the bill | Decision calls, preprocessing, follow-up model calls and human review |
Do not turn a vendor demo into a head-to-head benchmark. The models need the same inputs and a measured outcome. If OpenAI exposes image context, test an image-bearing set separately; sending images to only one API would be a different task. For consequences beyond a reversible queue assignment, keep a person in the decision path until the threshold is supported by your own results.
Where does WorkerKit fit?
WorkerKit's current decision workers use Jev for the judgment and WorkerKit for the surrounding job: fetch permitted items, apply typed questions, route according to the answer, record the run and escalate below a configured confidence floor. The model call is one step in that system. The worker's app grants, schedule, budget and run receipt govern the work around it. A support inbox triager is one concrete workflow to examine before building a custom classifier.
OpenAI's Decisions API is not a WorkerKit decision-worker provider as of this announcement. No published Decisions API request contract is available here to implement and test against. If that changes, the useful comparison will be whether it supplies the answer shape and uncertainty signals the worker needs, not whether both products have “decision” in the name. Structured output is the start of an integration contract; source permissions, escalation and receipts still have to be designed.
What to check when OpenAI releases the API reference
The launch makes Decisions API worth watching. Before replacing a working Jev flow or promising an integration, check five concrete items in OpenAI's documentation: the request and response schema; how it represents uncertainty and no-match cases; whether question sets are evaluated independently; image limits and data handling; and pricing, limits and model version behavior. Run the same evaluation set only after both interfaces are specified. Until then, the defensible conclusion is narrower: OpenAI has announced a Luna-based API for finite-answer decisions, and Jev has a documented typed decision interface today.
FAQ
Is OpenAI Decisions API available now?
OpenAI says it is in limited preview as of September 29, 2026, with a broad release planned in the coming days. Check the DevDay recap and current developer documentation for an updated access path.
Does Decisions API return Jev-style probabilities?
OpenAI's launch recap does not specify probability or confidence fields. Jev's API schema does specify them for Choice and Score, and a probability for Noul. Do not assume the formats are interchangeable.
Can Decisions API analyze images?
OpenAI explicitly says developers can supply text or images as context. It has not published image input limits or a request example in the recap.
Is Decisions API just a GPT-6 Luna call?
OpenAI says it focuses Luna's intelligence on finite-answer questions. That wording does not reveal the endpoint, implementation or billing contract. Treat Decisions API as its own announced product until OpenAI publishes those details.
Is OpenAI Decisions API cheaper than Jev?
There is no verified Decisions API price to compare with TypeSafe's published Jev rate. Any numeric answer today would be guesswork. Compare full workflow cost once OpenAI publishes terms and you can run the same task on both.
Can I use Decisions API in a WorkerKit decision worker?
WorkerKit decision workers currently use Jev. There is no verified Decisions API integration or published request contract to implement against as of launch day. Browse decision model kits for the workflows available now.