What is Cloudflare Clef? A decision model for your workflows
Cloudflare Clef is an open-weight AI model that returns decisions and probabilities from text or images, rather than writing a reply. Released on 1 October 2026 alongside the smaller Clef-flash, it is available through Workers AI or as downloadable weights. If your product needs to classify a document, route a ticket or score an incoming request, it is a candidate for that decision step.
You define the possible answers. Clef assigns probabilities to them, and your application decides what happens next. A support workflow could ask which department owns a message and how urgent it is; a separate model or a person would still write the response.
What does Clef give your application?
It gives your code typed answers within a schema you supply. Cloudflare documents three question types: choice selects a named option, score rates against ordered levels, and noul returns the probability that a statement is true. You can ask up to 64 questions in one request.
Both hosted models have a 65,536-token context window and support vision. That makes a visual decision possible, such as whether a submitted receipt is legible. It does not turn the result into a written explanation. The model card describes a single pass that scores the allowed answers without generating free-form text.
Cloudflare says the API follows Jev's System One format. If you already use Jev, that gives you a practical starting point for a trial. Keep your existing labelled cases and compare the returned decisions before changing production traffic. Compatible request formats do not establish identical answers.
Which version should you try, and what does it cost?
Start by testing Clef-flash if response time matters most, and Clef if getting a classification right matters more. That follows Cloudflare's positioning; your own results should decide. The hosted choices and published input rates, checked on 2 October 2026, are:
| Model | Size | Per million input tokens |
|---|---|---|
| Clef | 27B parameters | $0.24 |
| Clef-flash | 9B parameters | $0.09 |
For a text-only routing job with 100,000 tickets and 2,000 total input tokens per ticket, the input volume is 100,000 × 2,000 = 200 million tokens. At those rates, input charges are 200 × $0.24 = $48 for Clef, or 200 × $0.09 = $18 for Clef-flash. This is an input-cost estimate before free allowances, plan charges and the rest of your application.
Use Workers AI's documented binding or REST API for a hosted trial. Downloadable Apache 2.0 weights offer another route, but running them means providing the hardware and operating the service yourself. Cloudflare's model card reports testing Clef on an H200 GPU; downloading weights does not establish that your existing server can run them.
What should you check before putting it in a workflow?
Check decision quality and response time on the same cases your application will see. Cloudflare reports median latencies of 209.3 milliseconds for Clef and 38.8 milliseconds for Clef-flash across its benchmark runs. Those are vendor measurements, not an end-to-end latency promise for your deployment.
Build a test set that includes the awkward cases: messages that belong to two departments, missing information, unfamiliar wording and inputs outside your categories. Add a review option where your workflow needs one. Measure how often the model sends work to the wrong place, not just how often it produces a valid field.
Decide what a wrong answer costs before choosing an automatic-action threshold. A suggested queue can be corrected cheaply; closing a customer case or blocking a request needs a different acceptance test. During the trial, log the proposed decision alongside the human decision and measure where they disagree. Keep the action in your application code so you can change the policy without rewriting the classifier.
Sources
- Michelle Chen, Alex Reneau and Kevin Flansburg, Cloudflare: Introducing Clef, 1 October 2026. Launch, positioning, vendor latency results and header screenshot.
- Cloudflare: Workers AI changelog, 1 October; Clef and Clef-flash documentation; pricing; Clef model card. Read 2 October 2026.
Related reading: What is Jev?, the decision-model introduction in this series.
To have a decision step tested and built into your product or workflow, tell Vesprr what you need to classify or route.