← All posts
This post is written in English.
Lots of production AI work isn't writing essays. It's deciding: is this ticket urgent, did this trace fail, should a human look at this? "Decision models" are built for that. You give them input and get back a typed answer with a probability, not generated text.
What's on offer
Cloudflare's Clef, with weights under Apache 2.0, is one example. Cloudflare reports median latency of 209 ms for Clef and 39 ms for Clef-flash, and says Clef trails rival Jev on some benchmarks such as When2Call. Those are the vendor's own numbers.
An independent test
Hamel Husain and Antaripa Saha tested several of these as cheap evaluators on 98 scored traces from an apartment-leasing agent:
| Model | F1 after tuning | Cents per 100 runs |
|---|---|---|
| OpenAI Decisions (GPT-6 Luna) | 0.87 | 3.72 |
| Perplexity Decider 1.1 27B | 0.79 | 0.82 |
Liquid d1 and Clef Flash fell below their 80% precision and recall minimums after tuning. It's one small dataset and one broad task, and the authors say results vary by task, so this isn't a ranking.
Their most useful finding is about method. The default 0.50 threshold often leaves performance on the table, and you should check calibration before treating a score as a confidence level. They also warn: "Be intentional, don't just use F1 score because this post did!"
My read
This looks like a real shift in how teams will build. A full LLM is overkill for a routing decision, and a small, fast classifier that says "72% likely" is easier to reason about in code.
Just remember that a probability from an uncalibrated model is only a score. If you act on "0.8" as if it means 80% of the time, test that against real outcomes first.
