System One models
What it is
System One models (also called decision models) are a class of AI models trained to output a typed decision — a choice from a predefined set, a score, or a boolean — with a calibrated confidence value, instead of generating free-form text token by token. Given a state and a schema you define in advance, the model returns a guaranteed-conformant structured answer directly, with no autoregressive generation and no parsing of open-ended language required downstream. The name borrows Daniel Kahneman's System 1/System 2 distinction: fast, intuitive judgment versus slow, deliberate reasoning.
Why it's emerging now
TypeSafe AI introduced the category alongside its first model, Jev, on September 15, 2026, trained with a method it calls Reinforcement Learning for Calibrated Decisions (RLCD) and pitched as 20–200x faster and 40–400x cheaper than LLMs for classification, routing, and judge-style tasks inside agent pipelines — the high-volume, structured-decision calls that don't need prose. Within about a week, a second, independently built model — Contrastive-LM's open CLM-8B, trained with a different contrastive objective — shipped behind a TypeSafe-compatible API and was marketed explicitly as "an open System One model," turning what started as one company's product name into a category other builders are targeting. Coverage has been heavily critical as well as descriptive: several commentators argue the category is really a rebrand of existing classifier/encoder techniques, which is itself a sign the term is being argued over as a category, not just repeated as PR copy.
How it relates to what you know
It sits opposite reasoning models / test-time compute on your ledger: where reasoning models spend more inference-time computation to think longer, System One models are built to skip generation entirely for the decision-shaped subset of tasks. It also sharpens structured output (JSON mode, constrained decoding) — instead of constraining an autoregressive model's output to fit a schema, a System One model's architecture is schema-native from training onward, which is the source of the speed claims.