What is a decision model? AI decision models explained

In short

A decision model is an AI model that reads content you supply, answers typed questions about it, and returns calibrated probabilities instead of prose. You list the possible answers; it tells you which one fits and how likely each of the others was. Jev is a decision model, the first System One model from TypeSafe AI, whose launch post described the class as models “built to make fast, structured decisions that software can use directly”. Within weeks Cloudflare, OpenAI, Microsoft, Perplexity and Liquid AI had shipped their own, and most of them call the category decision models. System One model is TypeSafe’s name for the class and the umbrella term this site uses.

What goes in and what comes back

Every call has two parts. The state is the content the model reads: a support ticket, a product review, a passage from a document, or a photo for the models that accept images. The questions are typed, meaning each one declares the shape of its answer before the model sees the content. TypeSafe’s docs draw the line plainly: System One models “do not write replies, produce code, or generate explanations of their reasoning”.

There are three question types, and the names TypeSafe gave them have mostly stuck.

• Choice picks one option from a list you write. Jev accepts up to 255 options per question, and the answer carries a probability for every option, not only the winner.
• Score places the content on an ordered scale you define, up to 10 levels. The returned score is the probability-weighted mean of the level numbers, so it can land between levels.
• Noul asks a yes/no question and returns one number from 0 to 1: the probability that the answer is yes.

Choice and Score answers also carry a confidence value. It is not a second opinion from the model. TypeSafe calls it “a statistic computed from the probability distribution the answer already gives you”: high when the probability piles onto one answer, low when it is spread out. Your code picks the threshold.

Calibrated is the word that matters in that definition. A calibrated probability matches outcome rates: of all the answers a model gives at 0.8, about 80 percent should turn out right. TypeSafe says System One models are “trained for calibrated decisions: their probabilities are optimized against outcomes”, with the caveat that calibration “is measured across groups of predictions; it does not guarantee that an individual answer is correct”.

Other vendors keep the shape and sometimes rename the parts. OpenAI’s Decisions API has a predicate type for yes/no questions, with choice and score beside it. Databricks’ ai_decide SQL function keeps TypeSafe’s names, answering noul, choice and score questions against a table column. Liquid AI’s docs say that “every question you ask a decision model uses one of three types”.

Why the name: System One, decision model and the vendors’ labels

TypeSafe coined “System One model”. Its launch post credits Daniel Kahneman’s Thinking, Fast and Slow and the split between “fast, intuitive System 1 thinking and slow, deliberate System 2 reasoning”. Sorting a ticket into a queue is a System 1 task for a person. Writing the reply is not. The System One guide covers the name’s origin and what counts as a System One task.

The market reached for a plainer phrase. Cloudflare’s Clef post refers to “decision models such as Typesafe AI’s Jev System One model” and calls Clef “our open-source decision models”. Microsoft’s announcement opens with “Decision models are quickly emerging as an important new category in AI”. Nace.AI introduced Drex as “a small model that decides instead of writing”. Hanzo’s post says “Kai is Hanzo’s decision model”. Liquid AI files d1 under “Decision Models” in its docs, and Celeris calls Celeris-1 Decision a model “built for typed decisions”.

The product names vary more than the category name does:

• OpenAI: the Decisions API, which “evaluates text, images, or both and returns typed answers”.
• Perplexity: pplx-decider, which its docs call “a decision model”, served through Perplexity’s own Decisions API.
• Microsoft: Microsoft-Decision-1, a “decision-scoring model” that “returns a calibrated probability for each option instead of generated text”.
• Liquid AI: d1 as a hosted model, plus open d1-3B and d1-omni-600M to run locally.
• Databricks: ai_decide, a SQL function in beta that for each question “returns either a probability, a choice from named criteria, or a score on an ordered scale”.

So decision model is the thing, System One model is TypeSafe’s name for the class and this site’s, and classification model describes part of the job.

Decision model vs LLM

An LLM (large language model) predicts the next token and keeps going until it has written an answer. Ask it to classify a ticket and it still runs a generation loop to produce the word “billing”, which you then parse and validate. A decision model skips the loop. The output space is the set of options you wrote down, so the answer is a typed value your code can branch on, with a probability for each alternative.

Nothing is generated token by token, so the answer comes back sooner; OpenAI’s own framing for its Decisions API is “about 10x faster than the Responses API”. You pay for the state you send rather than for an explanation you did not ask for. And the output is a calibrated distribution rather than sampled text, which is what makes a threshold mean something.

What you give up is prose. A decision model cannot draft the reply, summarise the document or explain itself. The usual pattern keeps both: the decision model routes, filters and gates inside your code, and an LLM writes anything a person will read. The System One vs LLM guide has the side-by-side on latency, cost and failure modes.

Decision model vs classifier

A classifier sorts input into categories it learned at training time. A decision model does that job too, which is why “classification model” is a fair description of the Choice type. The differences are in the interface. You pass the labels with each request, so adding a category is an edit to a list rather than a fine-tuning run. The same model answers Score and Noul questions, which a label classifier does not. And the probabilities are trained to be calibrated, which a softmax over logits is not by default.

Whether that earns a new name was the first argument people had about Jev. Is Jev just a classifier? sets out both sides. The output shape is old; the per-request labels and the calibration target are what is new. The zero-shot classification model entry explains where the two overlap.

Which decision models exist

The models directory tracks every decision model with a public API or public weights. As of 12 October 2026 it lists the ones below. “Hosted API” means you call the vendor’s endpoint. “Open weights” means you download the model and run it yourself.

Some hosted models publish weights as well: Cloudflare says it is open-sourcing Clef on Hugging Face under Apache 2.0, and Liquid AI lists d1-3B and d1-omni-600M beside the hosted d1. The Jev alternatives guide compares them with the benchmark numbers behind each.

How to start

Pick one decision your code already makes badly with a regex or an LLM prompt, such as which queue a ticket belongs in or how urgent it is. Write the state as the model needs to see it and the question as a Choice, Score or Noul. Call a hosted model, log the probabilities, and set your threshold from the logs. The build guide covers state design, question wording and thresholds, and the recipes are worked examples with code in Python and TypeScript.

Where it does not fit

• Anything that ends in text. Replies, summaries, code and explanations are not what the model produces.
• Open-ended answers. If you cannot list the options in advance there is nothing to pick from.
• Arithmetic, counting and dates. TypeSafe’s jaggedness page says Jev 1.13 “does not count reliably” and “reads dates as text, not as ordered quantities”, and tells you to keep the arithmetic in code. It also warns against “System Two tasks: more layers of indirections”, meaning questions that chain several facts together.
• Decisions where one answer must be right. Calibration holds across groups of predictions, not per answer. Gate on confidence and send the low-confidence cases to a person or an LLM.

FAQ

What are decision models?

Decision models are AI models that answer typed questions about content you supply and return calibrated probabilities instead of generated text. You define the options; the model returns which one fits and a probability for each. They are used to route, filter, rank, gate and score inside software, with the decision itself as the output.

Is a decision model the same as a System One model?

Yes, in current use. System One model is TypeSafe AI’s name for the class, borrowed from Kahneman’s fast System 1 thinking, and Jev was the first one. Decision model is the plainer term most other vendors use, including Cloudflare, Microsoft, Liquid AI and Nace.AI. This site uses System One model as its umbrella term for the category.

Are decision models just classifiers?

They do a classifier’s job and more. The labels arrive with each request rather than at training time, the same model answers scale and yes/no questions, and the probabilities are trained to match outcome rates. Whether that earns a new name is a fair argument, and the is-Jev-just-a-classifier guide sets out both sides.

Should I use a decision model or an LLM?

Use a decision model for a judgment with a fixed set of answers that your code acts on: which queue, is it relevant, how severe. Use an LLM for anything a person reads. Most systems use both, with the decision model routing and gating inside the code and the LLM writing the text.

Can a decision model hallucinate?

Not in free text, because it never writes any. Its answer is always one of the options you listed. It can still choose wrong, and a probability can be off for an individual answer even when the model is calibrated overall. Treat confidence as a signal to gate on, not as a guarantee.

Which decision models can I use today?

As of 12 October 2026 the hosted options include Jev, the OpenAI Decisions API, Microsoft-Decision-1, pplx-decider, Clef, d1, Drex, Kai, Celeris-1 Decision, Mercury Decide, Decider 1, Solar Decide, Span-01 and Tev1. Kev, Laya, CLM and GLiNER2.5-Decide are open weights you run yourself. The models directory lists each one.