Decision models, explained: AI that answers with probabilities, not prose

Cloudflare and the AWS-backed Strands Agents project each published a "decision model" on 1 October 2026, and Perplexity documents a Decisions API of the same kind. All three describe a model that does not write prose: it scores a fixed set of answers you supply and hands back probabilities. This guide explains what that means and when you might use one.

What a decision model is

Perplexity's documentation gives the plainest definition: a decision model "reads natural-language text and images the way a multimodal language model does, but instead of writing text it returns typed answers with probabilities: yes or no, one of your options, or a level on your rubric." It adds that the model "does not write replies, generate code, or explain its reasoning; your code does the reasoning with the numbers it returns."

Cloudflare and Perplexity document the same three question types, and Strands describes similar ones:

• Yes/no: the probability that the answer is yes, from 0 to 1.
• Choice: a probability for every option you define, plus the most likely one.
• Score: a probability for each level of an ordered rubric (for example "Cosmetic", "Inconvenient", "Product unusable").

Cloudflare's example is a support message: you ask whether it is urgent and which team should handle it, and the model returns typed answers with probabilities "which your code can use to route the ticket, trigger an escalation, or defer to a human."

Both Cloudflare and Strands cite TypeSafe AI's Jev model, released in recent weeks, in explaining the current attention. Cloudflare says its Clef models are "fully Jev-API compatible".

When to use one instead of a general LLM

Strands sets out the trade-off directly. In exchange for less flexibility, it says, decision models "are faster and more capable at a given size, always produce an answer from the selected options, and can run with very low latency." The flip side, in its words, is that the approach makes them "significantly worse at solving complex problems than reasoning models", and unable to generate text, so they are "unsuited for coding, chatbots, document summarization, and other common LLM tasks."

Perplexity's guidance is to use one "where you would otherwise ask a chat model for a label and parse the reply: classifying, routing, grading against a rubric, or any decision you want to threshold." Strands describes hybrid agents that use LLMs "to make the hardest decisions" and decision models for "the easier rote decisions".

Strands' worked example shows the pattern: before an agent calls a weather tool, the 2B model answers two yes/no questions about whether the arguments were grounded in what the user said and whether the call is premature, so the agent asks which city was meant instead of guessing.

What a calibrated probability means

Strands defines calibration as "how trustworthy its confidence scores are", and measures it with the Brier score. Cloudflare says it trained Clef with a Brier loss "to refine probability calibration".

In plain terms (our gloss, not a vendor's): a model is well calibrated if, across many answers it scores at 0.8, about eight in ten turn out right. That is what makes a threshold meaningful. Strands argues this kind of reliability score "is not available through frontier LLM inference APIs".

Perplexity's documentation adds two practical cautions: values near 0.5 mean the model is unsure, and because identical requests can differ "in the second decimal place", you should "set thresholds with some margin".

The launches

Cloudflare Clef and Clef-flash (1 October 2026). Two models hosted on Workers AI, with weights on Hugging Face. The Hugging Face metadata lists the licence as apache-2.0, and the card says Clef is post-trained from Qwen3.8-27B. Cloudflare says Clef "is currently the leader when evaluated against the Jev Decision Index". These are Cloudflare's own results: its model card says they come from "our internal run" of the Decision Index 0.2.1 suite. That card table has 41 benchmark rows across six models; by its own bold marks, Clef-flash is best on 16, Clef on 11, Jev on 11, Kev 9B on 2 and DiffusionGemma Jev on 1. Jev is marked best on GPQA Diamond, MMLU-Pro and BBH. On latency, Cloudflare reports a median of 38.8 ms for Clef-flash and 209.3 ms for Clef, against 5.8 ms for Laya, which it says "trades off quality". Cloudflare is also offering fine-tuning of Clef, first through its forward-deployed engineers and "later as a self-serve fine-tuning platform".

Strands Decider 2B (1 October 2026). A 2-billion-parameter model built on Qwen3.5-2B, released with its training data and scripts; GitHub lists the repository licence as Apache 2.0, as does the Hugging Face metadata. The team reports a median of "around 115ms" on an Nvidia RTX 3090 and ranks it "3rd of 33 in the 2B class" on JevBench's public set, by its own measurement.

Perplexity Decisions API. One model, pplx-decider-v1-27b, priced at $0.04 per million input tokens with output tokens free. The page gives no launch date; it cites Perplexity's own timing tests on 30 September 2026.

Why it matters

On our reading, decision models are less a rival to chat models than a cheaper component beside them: Strands and Perplexity both say they do not write text, and Strands says they are significantly worse than reasoning models at complex problems. Every accuracy and speed figure above is the vendor's own, so the useful test is your own labelled data, and whether the probabilities hold up when you set a threshold on them.