Decisions API vs Jev in 30 seconds
- Same job. Both return one answer from a set you define, for classification, routing and agent tool choice.
- OpenAI’s Decisions API runs on a version of GPT-6 Luna, accepts images, and is in limited preview with no public pricing or docs yet.
- Jev is a non-autoregressive “System One” model, text-only, with public docs and published pricing of $0.042 per million input tokens.
- Early tests put them close on accuracy. Each was faster on different task types.
Side by side
| OpenAI Decisions API | TypeSafe Jev | |
|---|---|---|
| Company | OpenAI | TypeSafe |
| Announced | September 29, 2026 (DevDay) | September 2026 |
| Status | Limited preview; broad release “in the coming days” | Early access |
| Model | A version of GPT-6 Luna | Jev, first “System One” model |
| How it generates | LLM tuned for constrained decisions | Non-autoregressive, one parallel pass |
| Input | Text and images | Text, JSON, string arrays |
| Stated latency | “Less than a few hundred milliseconds end to end” | 70–500 ms end to end |
| Per-option probabilities | Not confirmed | Yes |
| Question types | Not documented yet | choice, score, yes/no (“noul”) |
| Pricing | Not announced | $0.042 / 1M input tokens, output free |
| Public API docs | Not yet | Yes |
How they work
OpenAI Decisions API
OpenAI describes the Decisions API as focusing “Luna’s intelligence on a specific set of user-defined questions with finite pre-defined answers.” You supply text or images and get back answers to classify content, route requests or choose an agent’s next action. According to The Decoder, OpenAI says it decides about ten times faster than GPT-6 Luna through the regular API, around 150 ms against 1.6 seconds.
TypeSafe Jev
Jev takes unstructured program state and returns typed, probabilistic decisions in a single parallel pass instead of generating text token by token. Because valid outputs are defined by the schema up front, TypeSafe says Jev cannot produce an answer outside it. TypeSafe calls this class of model “System One,” after the fast, intuitive mode of thinking.
Speed and accuracy in early tests
The most detailed public test so far is Every’s vibe check:
| Task | OpenAI Decisions API | Jev |
|---|---|---|
| Text tasks, correct steps (of 78) | 76 | 73 |
| Text tasks, typical latency | ~230 ms | ~500 ms |
| Classification accuracy | Effectively tied | |
| Classification latency | ~309 ms | ~161 ms |
One test on one set of tasks is not a benchmark. Latency also depends on where you call from: developers on Hacker News reported different numbers from Europe and the US. Measure on your own data before you commit.
API shape
Jev is documented: you POST a state and a map of named
questions to /v1/systemone. Question types are choice (options with
plain-language criteria), score (an ordered list) and noul (yes/no with a probability).
Answers include the choice, a confidence and per-option probabilities, and many questions can go in one call.
OpenAI’s Decisions API has no public reference yet. From the announcement, it follows the same pattern: you define questions and possible answers, send context, and get an answer back.
The decisionsapi.cc Decisions API docs use the same shape: context plus named
questions of type choice, score or yes_no, with probabilities
for every option.
Pricing
Jev lists $0.042 per million input tokens with output free. OpenAI has not announced Decisions API pricing; some coverage assumes standard GPT-6 Luna rates, which OpenAI has not confirmed. Every called pricing the factor that decides whether the Decisions API becomes one of DevDay’s biggest releases. Compare costs with the calculator →
Which to choose
- You need image input, for example classifying screenshots, documents or product photos: OpenAI’s Decisions API, once you have preview access.
- You need to ship this week on text, with docs and a known price: Jev.
- You want per-option probabilities for thresholds and audit logs: Jev documents them today; OpenAI has not confirmed them.
- You worry about depending on one vendor: wrap the call behind one function, and keep a labeled test set so you can compare providers, including an independent Decisions API.
What developers are saying
- “Jev killer.” Much of the launch-day reaction on X framed OpenAI’s Decisions API as a direct response to Jev, with image input as the headline difference.
- Price decides it. Developers repeatedly said they would compare cost per million calls before switching anything.
- Is this new? Skeptics on Hacker News pointed out that “people have been using LLMs as classifiers or rankers in a similar way for ages.”
- Lock-in. Others preferred open-weight models, warning that with a single-provider API, “if your access is revoked you’re screwed.”
Sources: OpenAI DevDay 2026 recap, The Decoder, Every, TypeSafe blog, Hacker News.
Decisions API vs Jev FAQ
Is OpenAI’s Decisions API a Jev competitor?
Yes. Both take context plus a fixed set of possible answers and return one of them. Press coverage of DevDay 2026 widely described the Decisions API as OpenAI’s answer to Jev.
Is the Decisions API faster than Jev?
It depends on the task. In Every’s early hands-on test, the Decisions API was faster on multi-step text tasks (about 230 ms vs 500 ms), while Jev was faster on classification (about 161 ms vs 309 ms). Both are early and numbers will move.
Does Jev support images?
Not at the time of writing. Jev accepts text, JSON and arrays of strings. Image input is the most visible difference in favor of OpenAI’s Decisions API.
Which is cheaper, the Decisions API or Jev?
Jev is the only one with published pricing: $0.042 per million input tokens and free output. OpenAI has not announced Decisions API pricing yet.
Can I switch between them later?
The request shapes are similar: context in, named questions with allowed answers, typed answers out. Keep the call behind one function in your code and switching is a small change.