the short answer
OpenAI announced its Decisions API at DevDay 2026, and Vercel says the GPT-6 Luna-powered service is in limited preview for questions over text or images. Jev already publishes its Choice, Score and Noul request contract and probability outputs. On October 2, OpenAI had not published a Decisions API schema, price or comparable benchmark. You can plan a test now, but you cannot make a measured speed or quality comparison yet.
- Shared job
- Choose from predefined answers for a bounded decision
- Jev contract
- Published Choice, Score and Noul questions with probabilities
- OpenAI contract
- Announced Decisions API; public request schema not verified October 2, 2026
- OpenAI model
- GPT-6 Luna, according to Vercel’s report of the announcement
- Access
- Vercel reports limited preview; verify with OpenAI before implementation
The API Details Still Missing
A Jev request sends shared state and typed questions, then gets probabilities for the allowed answers. You can read that contract today. OpenAI calls its new product a Decisions API, but its GPT-6 Luna model page describes a general model with structured output support. That page is not a Decisions API specification.
OpenAI named the API in its DevDay recap. Vercel reported text and image context, caller-defined answers and limited preview access. On October 2, 2026, the public sources still gave no Decisions API endpoint, request fields, probability definition, price or limit schedule. An integration plan has to leave those boxes blank for now.
Compare the Same Support-Routing Decision
Take the invoice ticket where the customer cannot sign in to download it. Under a "first blocker" rule, the expected route is account access, even though the message mentions billing. That example gives both systems the same job. Add tickets with two plausible owners and tickets missing a key fact. The table sketches a test, since OpenAI has yet to publish a request body.
| Evaluation field | Jev | OpenAI Decisions API |
|---|---|---|
| Evidence | One customer message and approved routing criteria in state | The same customer message and approved routing criteria as context |
| Question | Choice: which team handles this request first? | Question with the same predefined team answers |
| Allowed answers | Account access, billing, technical support, needs review | The same four answers, if the preview contract permits |
| Native result to retain | Selected Choice, full option distribution and model ID | Whatever answer and metadata the published preview returns; do not assume probabilities |
| Application action | Code routes or queues review | Code routes or queues review |
What Is Known, and What Must Be Measured
| Dimension | Current evidence | Fair test |
|---|---|---|
| Output shape | Jev documents typed probabilities; OpenAI preview schema is not public | Capture native output and parsing failures before normalization |
| Input modalities | Jev model card describes text; Vercel reports text and image for Decisions API | Compare text-only first; evaluate image tasks separately |
| Accuracy | No controlled cross-product result verified | Use blind labels and class-specific errors on the same ticket set |
| Calibration | Jev exposes distributions; OpenAI probability output is unverified | Only compute calibration for a defined, available probability signal |
| Latency and cost | No comparable end-to-end measurement | Include provider, retries, timeouts and review time at matched coverage |
How to Verify the Comparison on Your Data
Keep Jev thresholds with Jev. If OpenAI returns no comparable probability, you can still count routing errors and cases sent to review. There is no reason to manufacture a confidence score. The benchmark method covers paired labels and failures, while Jev limitations gives you useful stress cases.
- Write the routing rubric and collect reviewed tickets, including ambiguous and out-of-scope examples.
- Freeze train, calibration and untouched test splits by customer or conversation.
- Pin Jev version, question and state projection; record the OpenAI preview model and contract when access is granted.
- Run both on identical evidence with their native interfaces and preserve every response or failure.
- Compare errors at the same review coverage, then latency, total cost and failure behavior.
FAQ
Is OpenAI Decisions API the same as Jev?
No. They address bounded decisions, but Jev is TypeSafe AI’s model with a published typed contract; OpenAI announced a separate API powered by GPT-6 Luna.
Can I call OpenAI Decisions API today?
Vercel reported limited preview after DevDay. Confirm access and the current request contract in official OpenAI documentation before writing an integration.
Does OpenAI Decisions API return calibrated probabilities?
The public sources checked October 2, 2026 did not establish its probability fields or calibration. Do not assume Jev’s output semantics apply.
Sources
Checked against the sources below on October 2, 2026. Model versions, prices and limits change.
- OpenAI: DevDay 2026 recap
- OpenAI API: GPT-6 Luna model
- OpenAI API: Structured Outputs
- Vercel: What is OpenAI’s Decisions API?
- TypeSafe AI: Jev primitives
- TypeSafe AI: Jev 1.13 jaggedness
Change note: First comparison after the DevDay announcement; public contract still pending.