Clef vs Jev: Jev Wins on Accuracy, Clef-flash on Price

Jev, TypeSafe's decision model, is more accurate than Cloudflare's Clef. Clef-flash, the smaller Clef, is cheapest since a 9 October 2026 price cut. We re-ran the Clef models that day on DecideBench, our test of 400 decisions such as whether a refund is within 30 days. Jev got 98.0% right at $32.26 per million decisions, Clef 94.0% at $130.85, and Clef-flash 85.2% at $20.72.

Do you need the fewest wrong answers? Use Jev. It got 8 of 400 wrong, against 24 for Clef.

Do you already build on Cloudflare Workers, or need the model to read images? Use Clef.

Do you need to run the model on your own hardware? Both Clef models are free to download. Jev is not.

Do you make huge volumes of easy decisions, such as sorting support messages? Clef-flash is the cheapest of the three. Keep it away from checking claims and ruling on returns.

• All three are decision models. You send text, a question and a short list of options, such as "approve this refund or not", and the model picks one.
• We ran all three the same way, through the same gateway and with worked examples: one solved example for each answer option, sent with every question. The costs compare like for like.
• Cloudflare released both Clef models on 1 October 2026.
• Clef costs about four times as much as Jev per decision. Clef-flash costs about a third less than Jev.
• Both new models answered faster than Jev on our test.
• Clef takes Jev's request format, so switching is mostly a change of URL and key.

What changed on 9 October?

Cloudflare's Workers AI pricing page was last updated on 9 October 2026. It lists Clef-flash at $0.038 per million input tokens, down from $0.09. The larger model stays at $0.24, and neither charges for output.

Cloudflare also made both models faster that week, so we re-ran them the same day. Our changelog records the new numbers:

The speed change is real. The small model's slowest requests used to take over 12 seconds, and now they take under one.

The accuracy change is not. The larger model's score moved by 3 items out of 400, which is within the normal wobble between two runs. We had published it at 94.8% in our Yev post. Use 94.0% from now on.

Which is more accurate, Clef or Jev?

Jev is more accurate, with 98.0% right against 94.0% for Clef and 85.2% for the small model.

DecideBench is our own benchmark, so weigh it as such. Its 400 decisions come in 200 pairs. The two halves of a pair differ by one detail that flips the right answer, such as a return requested on day 30 or day 31 of a 30-day window. A model that guesses from topic words gets one half wrong. Our guide on how to test a classifier explains why we built it that way.

The pair column is the harder test. Out of 200 pairs, Jev got both halves right in 192. The larger model managed 176, and the small one 144.

The gap is not even across kinds of work. Each family has 50 decisions:

The larger model matches Jev on routing and on agent actions. It falls furthest behind on returns, where the answer turns on a date or a clause.

The small model is weakest at checking claims, with 66% right. Returns come next at 74%. Both tasks ask the model to hold a fact against a rule. On sentiment it matched its larger sibling.

With 50 decisions per family, one decision is 2 points, so read single-family gaps as directional.

What does Cloudflare's own benchmark say?

Cloudflare's launch post reports Clef ahead of Jev on most of its tests. These are the vendor's own numbers, and we have not reproduced them.

The two benchmarks disagree, which is common when a vendor picks the tests. Our pairs reward reading the one detail that flips an answer. Cloudflare's suite leans toward sorting intents and calling tools. Trust the test that looks most like your own work.

Which is cheaper?

Clef-flash is the cheapest, at $20.72 per million decisions. Jev costs $32.26, and the larger model $130.85.

Costs use each model's list price for the tokens billed. With worked examples, each test decision comes to about 545 input tokens on the two Clef models. At $0.24 per million tokens, the larger model's price per token makes it the dearest of the three. Our Jev pricing post breaks down the other side.

Take a support queue that makes 300,000 decisions a month at the error rates above:

At this scale the bill is small for all three. The people who check the wrong answers cost more than the model, so accuracy usually decides it.

Which is faster?

Clef-flash answered fastest, at a 417 ms median. The larger model took 484 ms and Jev 639 ms.

We called all three through AI Space, our gateway, from one client machine with 4 requests in flight. The gateway and the network sit inside every one of those numbers. They compare with each other, but not with figures measured inside a data centre.

Cloudflare reports much lower medians of its own: 209.3 ms for the larger model, 38.8 ms for the small one and 524.1 ms for Jev. Those leave out our network, and both sets agree on the order.

Can you run Clef yourself?

Yes. Both models are open weights, which means you can download them and run them on your own hardware. Cloudflare published them on Hugging Face under the Apache 2.0 licence, which allows commercial use. Jev is a hosted API only.

The hosted versions document image input only, not video. The model cards show how to serve either model yourself with a Jev-style /v1/systemone endpoint.

We did not self-host either model, so every number in this post comes from the hosted service. A copy on your own graphics card would have its own cost and speed.

For an open model that costs less to run, see Yev. Our roundup of Jev alternatives covers the rest.

Is Clef a drop-in replacement for Jev?

Mostly. Cloudflare calls Clef "fully Jev-API compatible", and our test sent all three models the same request in Jev's format.

On Workers AI the URL and the key change. You call @cf/cloudflare/clef through the REST API or the env.AI.run binding in a Worker, with the same state and questions fields.

Three differences to plan for:

• Images. The hosted model takes up to 4 images per request, as an extension to Jev's format. It does not accept remote image links.
• Length. The hosted larger model reads 65,536 tokens at once, about twice what Cloudflare's post gives Jev. The hosted small model reads less than Jev.
• Accuracy per task. Run your own decisions through both before switching, especially anything that turns on a date or a rule.

Cloudflare's launch post says it does not read, store or train on hosted requests unless you opt into fine-tuning.

Which should you pick?

Use Jev when accuracy matters most. Use Clef when you live on Cloudflare or need images. Use Clef-flash for cheap, easy, high-volume decisions.

A common setup uses two models. The cheap one answers when it is confident, and the rest go to the accurate one. Our LLM cascade guide walks through it, and the small model in front of Jev fits that shape. For other options, read OpenAI Decisions API vs Jev or decision model vs LLM.

FAQ

Is Clef better than Jev?

Not on accuracy. On DecideBench, Jev got 98.0% of 400 decisions right and Clef got 94.0%, at about four times the cost per decision. Clef answered faster and reads images. Cloudflare's own tests put it ahead on most tasks, but those are vendor-reported and use different tests.

How much does Clef cost?

On Workers AI, Clef costs $0.24 per million input tokens, with no charge for output. On our test that came to $130.85 per million decisions, against $32.26 for Jev. The smaller model got cheaper on 9 October 2026. It now costs $0.038 per million tokens, down from $0.09.

Is Clef open source?

The weights are open. Cloudflare published both models on Hugging Face under the Apache 2.0 licence, which allows commercial use. You can download them and serve them yourself. The model cards were tested on one NVIDIA H200 graphics card and offer a Jev-style endpoint.

Is Clef-flash good enough for production?

For simple sorting, often yes. It got 94% of customer-intent decisions right and 92% of sentiment ratings. For checking a claim against evidence it got 66%, and for ruling on returns 74%. Test it on your own decisions, or put a more accurate model behind it for low-confidence answers.

Why did Clef's DecideBench score change?

We re-ran it on 9 October 2026 after Cloudflare's speed update. Its accuracy fell by 3 decisions out of 400, to 94.0%, which is normal variation between runs. Its median wait fell from 811 ms to 484 ms. Our changelog lists every change.

Sources

• Cloudflare, Introducing Clef: our open-source decision models, 1 October 2026
• Cloudflare, Workers AI pricing, last updated 9 October 2026
• Cloudflare, Clef on Workers AI, accessed 14 October 2026
• Cloudflare, Clef-flash on Workers AI, accessed 14 October 2026
• Cloudflare, Clef model card and Clef-flash model card, Hugging Face, accessed 14 October 2026
• XY Space, DecideBench dataset and results, v1.1, accessed 14 October 2026
• XY Space, DecideBench leaderboard, accessed 14 October 2026
• XY Space, DecideBench changelog, 9 October 2026
• TypeSafe, Introducing System One Models and Jev, 15 September 2026