Decision model
Also known as: non-autoregressive model, System One model, decision AI, classifier model
Standard LLMs generate an answer by producing one token at a time, which takes time and costs money even when the question is simple: should this ticket go to billing or support? A decision model skips all of that. You describe the state, define the allowed options, and the model returns a probability for each option in one pass over the input. No text to write, no JSON to parse, no chance of a malformed reply.
The category went from concept to crowded market in roughly three weeks in September and October 2026. TypeSafe AI's Jev launched September 15 and defined the vocabulary. Cloudflare followed with Clef and Clef-flash on October 1, Liquid AI released open-weight d1-3B on October 7, and Microsoft and Perplexity each shipped their own variants. Each model targets the same class of workload: routing, classification, guardrail checks, content moderation, and tool selection inside an agent loop.
For builders, the appeal is straightforward. Decision models are 10x to 200x cheaper per call than frontier LLMs on the tasks they handle, answer in tens of milliseconds rather than seconds, and return typed outputs that code can consume directly. The trade-off is that they only work when you know your option set in advance. Open-ended reasoning still needs a generative model. The emerging pattern is to put a decision model in the hot path for cheap, fast choices and call a larger LLM only when the situation genuinely requires it.