Three weeks ago there was one service of this kind. TypeSafe announced Jev on 15 September. Cloudflare released Clef and Clef-flash on 1 October, and OpenAI opened its Decisions API to all developers on 6 October. We have already tested Jev on our own work, so we read the documentation for all of them to see what has changed.
What a decision API does
You send the current situation and a set of questions, each with a fixed set of allowed answers. The service returns an answer to every question with a probability attached, and it writes no text. That suits the small judgments inside a larger system. Is this message a refund request? Which team should take it? How urgent is it? Your own code reads the answer and acts on it.
You send
State and questions
Text, JSON or other evidence, with the allowed answers for each question
You get back
A typed decision
Yes or no, a choice or a score, each with a probability
The question types match across the three. One asks whether a statement is true and returns a probability between 0 and 1. Another picks one option from a list you supply. The last places the input on an ordered scale. Jev and Clef call the first type a "noul", and OpenAI calls it a predicate.
What each one can read
The services differ most in what you are allowed to send.
| Measure | OpenAI Decisions | Cloudflare Clef | Jev |
|---|---|---|---|
| Opened | 6 Oct 2026, public beta | 1 Oct 2026 | 15 Sep 2026 |
| Reads | Text, inline images | Text, JSON, images, video | Text, JSON |
| Limits per request | 128 images | 64 questions, 4 images | 64,000 tokens |
| Models | GPT-6 Luna | Clef (27B), Clef-flash (9B) | Jev |
| Runs on | OpenAI's API | Workers AI, or your own hardware | TypeSafe's API |
OpenAI's endpoint accepts images only as data placed inside the request. It doesn't take links to images, files, audio or video. OpenAI's documentation describes pairing it with the Live API, so that a voice session can hand a question to Decisions, but Decisions itself never receives the audio.
Clef reads the widest range. Cloudflare's model page lists text, JSON, images and video. The launch note mentions only images, up to four per request, so we would test video before planning around it. Cloudflare says Clef uses the same request format as Jev, and that an existing Jev integration moves over by changing the endpoint and the model name. The weights are on Hugging Face under the Apache 2.0 licence, which means you can run the models yourself.
Jev reads text and structured data only.
Speed
Only one of the three companies has published a comparison. It is Cloudflare, testing its own models against Jev.
Cloudflare's benchmark · 43 runs · Vendor-reported
Clef-flash answers in 38.8 ms.
Median response time, with the p95 beside it. OpenAI has published no figure in milliseconds.
Median response time
In this test Clef-flash is 13 times faster than Jev at the median, and Clef is 2.5 times faster. The p95 is the time that 95 calls in 100 come in under. For Clef-flash it is about three times the median, 122.4 ms against 38.8 ms, so some calls are much slower than the typical one.
Read the Jev figure with care. Cloudflare doesn't say how it reached Jev. The model card gives a single H200 GPU for Clef's 209.3 ms, and an independent write-up from launch day describes the Jev number as a round trip to TypeSafe's hosted service. If that is right, Jev's time includes a trip across the internet and Clef's doesn't. The same write-up measured Clef-flash at 191 to 205 ms from Italy through Cloudflare's public API, about five times the published median. TypeSafe quotes 70 to 500 ms for Jev. In our own dental booking test on 27 September, a single Jev call took 806 ms measured from our side.
OpenAI says Decisions returns answers about 10 times faster than its own Responses API. That compares OpenAI with OpenAI, and it gives no time in milliseconds.
We would measure from our own servers, with our own requests, before quoting any of these numbers to a client.
Price
All four models charge for input tokens only. None of them bills for output.
Published list prices · Checked 7 October 2026
Jev lists lowest, at $0.042.
US dollars per million input tokens. Output is free on all four.
Input price per million tokens
These are small sums. One guide puts a typical three-question request at about 450 tokens. At that size, a million decisions cost about $19 on Jev, $41 on Clef-flash, $45 on OpenAI Decisions and $108 on Clef.
The list rate is only part of the bill. OpenAI's documentation says regional processing and long requests cost more. Images are counted as tokens too, and each service counts them in its own way, so a rate per million tokens says little about a job that is mostly pictures. Work out the cost of one real decision on each service and compare those.
Accuracy
We left accuracy out of the charts because the only figures we found come from Cloudflare. Its model card has Clef ahead of Jev on 24 of 42 measures and Jev ahead on 11. Jev's leads include two reasoning tests with wide gaps: 92.9% against 73.7% on BBH, and 78.3% against 48.0% on GPQA Diamond. We found no independent test that covers all three services.
Where we would start
| If the work is | Start with |
|---|---|
| Images or video, and speed comes first | Clef-flash |
| Images or video, and you want the larger model | Clef |
| Text only, at the lowest list price | Jev |
| Already built on OpenAI, or decisions inside a voice session | OpenAI Decisions |
The table is a shortlist. It doesn't name a winner, because the published numbers can't support one. Before you commit, replay a sample of your real requests through each candidate and compare four things. How often is the answer right? Can the probabilities be trusted, so that answers given at 90% are right about nine times in ten? What are the median and p95 response times from your own servers? What does one decision cost?
For our front-desk agents the input is a patient's message, which is text. If we were choosing today we would keep Jev, the option we have already tested, and put Clef-flash through the same dental booking test next. It accepts the same requests, and a speed claim that large is worth checking ourselves.
Sources
- OpenAI API docs: Decisions guide
- Mixed News, OpenAI's Decisions API runs on GPT-6 Luna alone (request limits)
- Cloudflare, Introducing Clef: our open-source decision models
- Cloudflare changelog, Clef on Workers AI (latency table, 43 benchmark runs)
- Cloudflare Workers AI: Clef model page
- Cloudflare Workers AI: Clef-flash model page
- Hugging Face, cloudflare/clef model card
- Flavio Copes, A deep dive into Clef (independent timings, prices)
- TypeSafe AI, Introducing System One Models and Jev
- OpenRouter, What is Jev? (inputs, request size, 450-token example)
- ADROLAB lab note, Jev vs GPT-6 Luna: how they compare (our own run, 27 September 2026)