Hi everyone. A support ticket arrives: “Checkout has been failing for every customer for the last hour.” Your system has to decide fast: is it urgent, is it billing or technical, who gets it.
One way is to ask a chatbot for a reply and then parse the text it writes. The other is to ask a model that never writes a word and only returns numbers. Clef, which Cloudflare released on October 1, 2026, is the second kind.
A note on how this article was made: I used Claude to help read the sources and draft the text. I checked the numbers against the sources linked at the end.
Every number below comes from Cloudflare, the Hugging Face model card, TypeSafe or one independent test, and I say who measured what.
A normal language model writes text one token at a time. A decision model does not write at all. You send it a state (a ticket, a web page, an agent log) and a set of typed questions, and it scores every allowed answer in a single pass.
Think of a multiple-choice answer sheet with a confidence level printed next to every box.
The docs define three question types:
Here is the request from Cloudflare’s docs for the ticket above (abridged):
curl https://api.cloudflare.com/client/v4/accounts/$CLOUDFLARE_ACCOUNT_ID/ai/run/@cf/cloudflare/clef \ -H "Authorization: Bearer $CLOUDFLARE_API_TOKEN" \ -d '{"model": "clef", "state": "Checkout has been failing for every customer for the last hour.", "questions": {"urgent": {"type": "noul", "instructions": "Is this support request urgent?"}}}'
Classmethod’s write-up shows this request returning 0.9917 for urgent, with 154 input tokens and 0 output tokens. There is nothing to parse, so your code can simply branch on the number.
There are three models:
The pricing page lists only an input price, because there is no output to bill. Clef’s weights are on Hugging Face under Apache 2.0.
The biggest difference is the weights. Clef’s are public under Apache 2.0 and Jev’s are not. That makes Clef open weights, not open source, since the model card says nothing about the training data or the recipe.
Clef also reads images, while Jev, according to Cloudflare, is text-only. The API is meant to be a drop-in: Cloudflare calls Clef fully compatible with Jev. Classmethod found one catch, though. Over REST, Clef wraps the answers in a result field, so code written for Jev needs a small change.
Then there is price. Jev costs $0.042 per million input tokens with free output (TypeSafe’s own figure). Next to that, Clef is about 5.7 times more expensive and Clef-flash is a little cheaper.
Cloudflare ran its benchmarks itself, on a set called the Jev Decision Index, and published the tables on the model card. It says Clef leads the index. A few of the rows, Clef against Jev:
The card gives no confidence intervals, so I would not call a gap of one or two points a win.
The big gaps on the knowledge tests, GPQA and MMLU-Pro, point the other way: a decision model is not a general reasoner. If your question needs hard reasoning, test before you switch.
Cloudflare’s own runs give a median latency of 209 ms for Clef, 39 ms for Clef-flash and 524 ms for Jev.
The one independent test I found, by Classmethod, made five calls of one scenario with three questions and got the opposite order: Jev 0.238 s, Clef-flash 0.294 s, Clef 0.478 s. The author says the sample is too small to conclude anything, and it ran before Clef got faster on October 9.
I would not trust either set on its own. One is Cloudflare’s run, the other is five calls from one person’s machine. Measure from your own region.
Some early coverage, including the Classmethod and DevelopersDigest pieces I read, quotes the old numbers. Here is what moved:
A quick calculation with list prices (my arithmetic): one million decisions with a 2,000-token state cost $480 on Clef, $76 on Clef-flash and $84 on Jev. Before October 9, Clef-flash would have cost $180.
A decision model works when the answer comes from a fixed list and you can check it afterwards. Typical jobs:
It is the wrong tool for chat, for writing text or code, and for open-ended reasoning. One more caution: the output is always a valid answer, but a valid answer can still be the wrong one.
By list price and Cloudflare’s own numbers, Clef-flash is the easiest one to try first: it is now cheaper than Jev, and its median is 39 ms. How it holds up on real tickets is still an open question.
What the sources support is narrower. Clef returns probabilities, not text, so it belongs in routing, triage and guards, not in chat. The benchmarks are Cloudflare’s own and the picture is mixed. Before you switch anything, check it on your own data.
Have you tried a decision model in a real workflow yet?
If you enjoyed this article, I’d be grateful for your support
I’m a Product Manager with an engineering background (ex SWE), focused on building, scaling and growing products. I’m especially interested in new technologies, particularly AI/ML, and how they can be applied in real workflows.
Email: akzhankalimatov@gmail.com
✅ Connect with me on LinkedIn, X (Twitter)
✅ Press and hold the 👏 to give up to 50 claps for this article
What is Clef? The Cloudflare Model that Competes with Jev was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.