Jev vs Clef: two decision models, compared on what has been measured

At a glance

Clef is the first decision model from someone other than TypeSafe that copies Jev’s contract outright: Cloudflare launched it on 2026-10-01 as “fully Jev-API compatible”. That makes the comparison unusually clean, because you can send the same request to both. Every number below says where it comes from: (vendor) is the company selling the model, (our measurement) is a call we made ourselves.

Credits are standard / large: an evaluation is large above 8,000 characters of state or five questions. The vendor latencies are Cloudflare’s numbers and are not comparable with each other — Clef was measured inside Cloudflare’s network and Jev across the public internet.

The difference that matters most: how much of the state is read

Cloudflare’s model page lists a 65,536-token context. In practice, on Workers AI, Clef reads far less of the state than that. We tested it on 2026-10-04 by hiding one decisive sentence — a customer asking to cancel — at different points in 20,000 characters of neutral text, and asking a single yes/no question about it:

Both models found the sentence anywhere in the first 12,000 characters and missed it entirely at the end. The token count Cloudflare billed stayed at 2,190 whatever we sent, which matches a cap of about 2,048 state tokens. Text past the cap is not read and not billed. The question set is not capped the same way: twelve long questions still counted as more than 8,000 tokens.

For English prose, 2,048 tokens is about 12,000 characters. For code, JSON or Chinese, Japanese and Korean text it is much less. Jev’s budget is about 32,000 tokens shared between the state and the longest question, so a long support thread, a contract or a full chat transcript fits. The full method and every number are in Cloudflare Clef, tested.

This may change — Cloudflare could lift the cap at any time — so check the date at the top of this page.

Benchmarks: what Cloudflare reports, and what nobody has checked

Cloudflare published results on its Jev Decision Index, 43 benchmarks across 10 metrics. These are vendor numbers, and at the time of writing no independent group has reproduced them.

Two cautions. First, do not mix these with older Jev numbers: the independent jev-benchmarks pilot measured Jev on Banking77 with plain accuracy on a 100-example sample, a different metric and a different sample from Cloudflare’s macro-F1. Second, the pattern is plausible on its face — Clef wins on fixed intent lists and tool lookup, Jev wins where the model has to decide whether to act at all — but that is a reading of vendor data, not a finding.

Price, honestly

Per token, Jev is the cheapest of the three. A typical evaluation — three questions and a short paragraph — cost us $0.000020 on Jev, $0.000033 on Clef-flash and $0.000089 on Clef at list price. At the other end, a request with 12 long questions costs up to $0.00038 on Jev and $0.0024 on Clef.

Cloudflare’s free allowance changes the picture for small volumes. Every account gets 10,000 Neurons a day, which at Clef’s rate covers about 458,000 input tokens: around 1,200 typical Clef evaluations or 3,300 Clef-flash evaluations a day. If you already deploy on Cloudflare and are happy to write the integration, calling Workers AI yourself is the cheapest way to run Clef, and How to get Clef shows how.

JevStation charges per evaluation in credits, not per token. Clef-flash costs the same as Jev, 1 credit for a standard evaluation; Clef costs 2, because its token price is almost six times Jev’s. What you get for that is both models behind one key, a playground, batch runs, MCP tools, saved threads and automatic refunds when an evaluation fails. Pricing has a calculator for each model.

When to pick which

• The input is long (a thread, a document, a transcript): Jev. Clef will judge only the first part, silently.
• You need images judged: Clef or Clef-flash. Jev does not take images.
• Short inputs, high volume, latency matters: start with Clef-flash. It is the fastest of the three and costs the same number of credits as Jev here.
• Fixed intent lists, tool lookup: try Clef; Cloudflare’s own numbers favour it there.
• “Should the agent act at all?” questions: try Jev; it led on When2Call.
• The data cannot leave your machines: self-host Clef. Jev has no open weights.
• Lowest cost per token at scale: Jev, if you can get TypeSafe access.

Whatever the table says, the deciding test is your own data. Paste twenty real inputs into the side-by-side runner, run them on both models at once, and compare the answers to what you know is right. For every other kind of alternative, including LLMs and trained classifiers, see Jev Alternatives and Jev vs open models.

Sources

• Launch, benchmarks, latency, data policy and fine-tuning. Cloudflare blog (vendor)
• Model IDs, limits and context window. Workers AI Clef model page and changelog (vendor)
• Prices and free Neurons. Workers AI pricing (vendor)
• Jev price and context. TypeSafe model docs (vendor)
• The 2,048-token plateau, first reported. Cloudflare community thread (independent)
• Truncation test, token counts and per-evaluation cost. Cloudflare Clef, tested and What One Jev Evaluation Costs (our measurement)

JevStation is independent. Jev and Clef are names of their respective owners, TypeSafe and Cloudflare; we are not affiliated with either.