What is a decision model? How AI that picks answers differs from chatbots

Decision Models Digest, Oct. 28, 2026

A decision model is an AI model that chooses an answer from a list you supply instead of writing text. You give it some input, a question and the allowed answers. It returns a probability for each answer in a single pass.

Where did decision models come from?

TypeSafe AI launched the first widely used decision model, Jev, on Sept. 15, 2026. TypeSafe calls it a "System One" model. The term comes from Daniel Kahneman's book "Thinking, Fast and Slow," where System One is fast, intuitive thinking. In that framing, chat models are the slower System Two.

Within five weeks, Perplexity, Cloudflare, AWS, OpenAI, Microsoft and Liquid AI had all released decision models of their own. Dozens of open-source versions followed.

How does a decision model work?

A developer defines the questions and the allowed answers. Most decision models support three kinds of question:

Yes/no questions, which return the probability that the answer is yes.

Choice questions, which return a probability for each option.

Score questions, which rate something on an ordered scale.

The model reads the input once and scores every option. It doesn't generate text, so its answer can't wander off format and doesn't need to be parsed. Many open versions start from an existing language model and replace the part that writes words with a small scoring layer. AWS's Strands Decider 2B, for example, swaps the text output of Alibaba's Qwen3.5-2B for a scoring head of about 1 million parameters.

What are decision models used for?

They're built for narrow decisions that software makes thousands of times a day. Examples include routing a support ticket, classifying a website, checking a claim against evidence, flagging a suspicious transaction and approving an AI agent's next action. They also serve as guardrails that check a chat model's output before it reaches a user.

Why not just use a chat model?

Speed and cost. Jev charges $0.042 per million input tokens and nothing for output. TypeSafe says a decision takes 70 to 500 milliseconds, though that's its own measurement.

Chat models can be more accurate. On DecideBench, an independent benchmark of 400 decisions, three general-purpose language models from DeepSeek and Zhipu scored 99.2% to 99.8%. The best decision model, Nace.AI's Drex Max, scored 98.8%. But the chat models cost 6 to 14 times more per task. The trade-off is a little accuracy for a large saving at scale.

What are the limits?

A decision model can only choose from the options it's given. If the right answer isn't on the list, it picks the closest one.

It can also be confidently wrong. Its probabilities are only useful if they're well calibrated, meaning an answer given 90% confidence is right about 90% of the time. JevBench measures calibration for that reason.

Quality varies widely. On DecideBench, the strongest models scored above 95%. Some small open models scored between 35% and 60%, against 22% for random guessing.

Finally, most speed and cost figures come from the companies selling the models. Independent benchmarks are still catching up.

How can I compare decision models?

The Models directory on this site lists every decision model we track, with makers, launch dates and scores from three independent benchmarks. You can sort by any benchmark or compare up to three models side by side.