How is a decision model different from an LLM?

An LLM generates text one token at a time and has to be asked to name an option. A decision model scores every option at once and returns the probabilities, so the answer is always one of your options and its confidence is part of the output. Decision models usually charge for input tokens only.

What are examples of decision models?

As of 20 Oct 2026: TypeSafe Jev (the first, launched 15 Sep 2026), Cloudflare Clef and Clef-Flash (open weights, Apache 2.0, on Workers AI since 1 Oct 2026), Perplexity pplx-decider, OpenAI's Decisions API (public beta), Microsoft-Decision-1, and open models such as Kev.

What is a System One model?

"System One" is a name used for models that answer a typed question in one forward pass instead of reasoning step by step in text. Decision models pick from options the caller lists; extraction models fill in typed values.

What is a decision model?

Q: What is a decision model?

A decision model takes a state (text, JSON, sometimes images) plus typed questions, and returns a probability for every option the caller listed, in one forward pass. It cannot answer with anything outside the list.

A decision model answers a question by choosing among options you list, and tells you how sure it is about each one. You give it a state, such as a support ticket, a pull request or a JSON record, and a question such as "which team should handle this?" with the teams as options. It returns a probability for every team, in one pass, without writing any text.

What goes in, what comes out

You send:

The state: text, JSON, and for some models images.
One or more questions, each of one type: choice (pick one of the options you list), score (a point on an ordered scale you define) or yes/no.
For each question, short instructions and the options or levels.

You get back, for each question:

A probability for every option or level, summing to 1, or for yes/no the probability of yes.
The top option and its confidence.
For score questions, the probability-weighted score.

TypeSafe, Cloudflare Workers AI and Microsoft Foundry use the same request format, which TypeSafe calls System One, so code written for one can switch to another by changing the endpoint and model name.

How it differs from asking an LLM

You can ask an LLM to pick an option and say how sure it is, but the confidence is text it wrote, and the answer can fall outside your list. A decision model scores the options directly, so the answer is always one of yours and the probabilities come from the model, not from a sentence. It answers in one forward pass, and the hosted ones charge for input tokens only.

That doesn't make decision models right by default. Whether their probabilities can be trusted, whether 90% sure really means right nine times in ten, is exactly what this benchmark measures.

What people use them for

Gating an agent's tool calls: approve, block, or ask a person.
Routing tickets and issues to the right team.
Moderation and triage.
Scoring against a rubric.

In each case the pattern is the same: act when the model is confident enough, hand off when it isn't. The threshold is only as good as the model's calibration.

Decision models as of 20 Oct 2026

Jev 1.14 and Jev 1.13, TypeSafe. The first decision model, launched 15 Sep 2026. $0.042 per million input tokens. Jev 1.14, available since 14 Oct 2026, adds image input behind an opt-in header.
Clef and Clef-Flash, Cloudflare. Open weights under Apache 2.0, released 1 Oct 2026, hosted on Workers AI at $0.24 and $0.038 per million input tokens. Both accept up to four images.
pplx-decider v1.2 and pplx-decider-flash v1.2, Perplexity. 27B and 4B, $0.02 and $0.01 per million input tokens, with image input.
OpenAI Decisions, in public beta, $0.10 per million input tokens, with image input.
Microsoft-Decision-1, on Microsoft Foundry.
Kev, Jared Palmer, open weights under Apache 2.0. Kev-4B is hosted on OpenRouter. Other open models such as Von, Laya, H2O-Lightning-4B and Chon-1.0-12B have no hosted API yet.

What vendors say, and what is known independently

Cloudflare's leaderboard of 1 Oct 2026 scores Clef ahead of Jev; Cloudflare built and ran it. TypeSafe publishes no benchmark table for Jev. The community-run JevBench and the Artificial Analysis Decision Index test decision models on written or academic items. As of 20 Oct 2026 we found no independent test against real recorded outcomes. Here is how we plan to run one.