Two decision models in one day

On 1 October, two of the largest companies on the internet released the same kind of model on the same day. Cloudflare published Clef and Clef-flash.1

AWS, through its Strands Labs team, published Strands Decider.2 Both are open, under the Apache 2.0 licence, and both are something most AI teams haven't used yet: a decision model.

What a decision model is

A language model writes. You ask it a question and it produces text, one token at a time. A decision model doesn't write at all. You give it the situation and a fixed set of answers, and it returns which answer it picks and how sure it is.

Cloudflare's own definition is a good one: a decision model "makes classifications to help agents decide how to act, based on certain probabilities." Its models return typed answers with probabilities that your code can use to route a ticket, trigger an escalation, or defer to a person.

The AWS model shows how simple the idea is underneath. Strands Labs took a small pretrained language model, removed the part that generates text, and replaced it with a small head, about a million parameters, that scores each option in one pass. No generation, no decoding loop. It answers three kinds of question: yes or no, pick one of several, and rate on a scale.

Most of what an AI agent does isn't writing. It's deciding: which tool, which team, whether to act or ask. These models are built for exactly that step.

The headline is speed

Because nothing is generated, these models are fast. Across 43 evaluations, Cloudflare reports a median latency of 209.3 ms for Clef and 38.8 ms for Clef-flash, against 524.1 ms for Jev, the decision model they compare with. AWS reports a median of 115 ms per question for its 1.9-billion-parameter model on a single RTX 3090, and says it also runs on an Apple-silicon Mac or a CPU.

Those are the companies' own numbers, measured their own way, so treat them as a starting point rather than a verdict. But the direction is clear. A decision that once needed a large model and a second of waiting can now run in tens of milliseconds, on your own hardware, for almost nothing.

The number that matters more

Speed is easy to measure and easy to sell. The number that decides whether you can trust these models is a different one: calibration. When the model says 0.9, is it right about 90% of the time?

The AWS team answers this directly, and it's the most useful line in either release. On short classification tasks the model hasn't seen, they report that answers given with a confidence of 0.9 or more are right about 95% of the time, and they add their own advice: below that, confirm the answer or ask a person.

That advice is the whole idea of the book in one sentence. A probability isn't an action. Someone still has to decide what happens at 0.9, at 0.6 and at 0.2. The answer isn't one line at 0.5. It's three doors: act on the confident cases, send the uncertain middle to a review, and escalate the cases that are both likely and costly. Machines take the easy ends; people take the middle.

What to do before you adopt one

Check calibration on your own data. A model calibrated on a benchmark can be overconfident on your tickets, your alerts or your documents. Label a few hundred real cases, group them by the model's confidence, and see whether 0.9 really means about nine in ten.

Draw the lines by cost, not by habit. Write down what a wrong "act" costs, what a review costs, and what a missed escalation costs. The thresholds follow from those numbers, and from how many reviews your team can actually do in a day.

Plan for more review work, not less. When decisions get cheap, teams make far more of them, and a rule set as a share of volume sends more cases to people. That was last week's letter, the Jevons paradox of decisions, and it applies here directly.

Keep the language model for what it's good at. The design these releases point to is a hybrid: a decision model makes the call quickly and reports how sure it is, and an LLM reads the messy input and writes the reply.

Two big companies shipping the same idea on the same day doesn't prove it works for you. It does mean the decision layer of agents is no longer a niche idea. It's becoming a normal part of how AI products get built, and the teams that check calibration and draw their lines by cost will get the most from it.