Same question, two engines. See what changes in question types, image input, refusal, price and latency — and convert your requests from one format to the other without rewriting anything.
OpenAI launched the Decisions API: you send a text or a photo with closed questions and get back only the answer — yes/no, one option from your list or a score — with its probability, no running text. It is the same idea as Jev, the decision model INEMA already studies. This project explains the differences with the source of each piece of information and ships a Python bridge that converts questions from one format to the other and compares the answers. To study, just read and run the offline demo; to call the real APIs, you need your own keys.
A decision model does not write: it chooses. Your system asks the question, defines the valid options and gets an answer ready to use in code — plus a probability so you know when to call a person.
Every line says where it came from: OpenAI's official documentation, the Jev project code, or what the authors of two videos measured. Nothing is presented as our own measurement.
Converts the request from one format to the other (noul ↔ predicate, criteria ↔ choices/levels) and normalizes both answers into a single format.
Three fictional cases with simulated answers run with no key and no cost; 12 tests guarantee the conversion. Live mode only runs when you authorize it.
Both answer the same three ways of asking. The differences are in the input, in refusal and in price.
| OpenAI Decisions API | Jev (TypeSafe) | |
|---|---|---|
| Yes or no | predicate → probability from 0 to 1 | noul → probability from 0 to 1 |
| One option from a list | choice + choices, com confidence | choice + criteria, com confidence |
| Score on a scale | score + levels (0-based index) | score + criteria as a list |
| Input | Text and image (base64 only) | Text only |
| Refusal | May answer refusal | Always picks among the options |
| Model | gpt-6-luna (only one) | ~typesafe/jev-latest |
| Price (input) | US$ 0.10 per 1M tokens, no output charge | US$ 0.042 per 1M tokens |
| Status | Public beta since Oct 6, 2026 | Available, no waitlist |
Sources: OpenAI's official guide (developers.openai.com/api/docs/guides/decisions) and the inematds/jev project. Details in docs/comparacao.md (in Portuguese).
Two creators tested the APIs right at launch. The numbers below are theirs, with few and easy questions — they are an indication, not a benchmark.
Jev came out about 58% cheaper per input token; in the incidents test, the total cost was about half.
Average of 146 ms (OpenAI) versus 155 ms (Jev); the median was slightly better on Jev. In an integration test, one call took ~300 ms, double the advertised figure.
A tie: both got everything right in the authors' tests. With photos, OpenAI told apart a damaged product, an intact one and a sealed box.
Sources: the video "OpenAI's Decisions API just dropped. Here's how it compares to Jev" and a second video, in Spanish, that integrated the API into a veterinary clinic triage (received only as a transcript).
The bridge translates the questions, calls each provider in its own format and returns the answers side by side, with a review alert when confidence is low or there was a refusal.
The bridge uses only the Python standard library. Keys are only needed in live mode.
No dependencies to install.
git clone https://github.com/inematds/decisions-jev.git cd decisions-jev
Only for live mode. Created on the OpenAI platform; billed separately from any subscription.
export OPENAI_API_KEY=...Only for live mode, to call Jev through OpenRouter.
export OPENROUTER_API_KEY=...Start offline. Only call the APIs once the question is well formulated.
Three fictional cases — pet on-call, a checkout incident and a refusal — with simulated answers. It shows the comparison format, with no key and no cost.
python3 -m ponte demo
The bridge turns noul into predicate, criteria into choices or levels, and each question's name into name.
python3 -m ponte converter jev-openai exemplos/plantao-pet.jev.json
The image becomes a base64 data URL inside the message. In the opposite direction, the bridge warns that Jev will receive only the text.
python3 -m ponte converter jev-openai exemplos/plantao-pet.jev.json --imagem foto.jpg python3 -m ponte converter openai-jev exemplos/devolucao-foto.openai.json # warns: 1 image discarded
Calls both providers with the same question and shows answers, latency and cost. It requires PERMITIR_API=1 so you do not spend by mistake. INEMA has not run this mode yet.
PERMITIR_API=1 python3 -m ponte ao-vivo exemplos/plantao-pet.jev.json
The ready-made prompt tells the agent to read the official Markdown documentation before coding, ask the questions in a single call, flag confidence below 0.6 as "review" and handle refusal without breaking.
# copy the text block from
docs/prompt-integracao.md
12 tests cover conversion in both directions, round trip, image, refusal, review and the live-mode block.
python3 -m unittest discover -s testes
For a person waiting, a few milliseconds make no difference. What matters is the input, the volume and what to do with doubt.
Decisions API. A product photo, a wound photo or a screenshot goes in the same call. Jev still reads only text.
Jev. The input price is less than half; across millions of records, the difference shows up on the bill.
OpenAI may refuse with refusal. With Jev, include an "insufficient" option. In both, a refusal or low confidence goes to a person.
What already exists and what is left to measure.