A plain-English explainer of the "System One Model" category in AI — the shape of the machine, where the name comes from, and how it differs from the LLMs you already know. Every claim is labeled: vendor claim = said by TypeSafe, unverified; observed = demonstrably true; analysis = this site's reasoning.
A System One Model is a new kind of AI model that makes fast, structured decisions instead of generating text. You give it a piece of unstructured state — an email, a support ticket, a transaction description — plus a list of possible options. It returns one calibrated probability per option, in a single pass, with no generated prose in between.
The first commercial example is Jev, announced by TypeSafe AI on Sep 15, 2026 observed. The headline numbers from that launch — roughly 100× the speed and efficiency of existing LLMs on decision tasks, and "cannot hallucinate" by design — are vendor claims, not yet independently verified as of 2026-09-20 vendor claim analysis.
Psychology's dual-process theory — popularized by Daniel Kahneman's Thinking, Fast and Slow — describes two modes of thought: System 1, fast, automatic, instinctive; and System 2, slow, effortful, deliberative observed.
The AI mapping flips it usefully. Large language models are System 2 machines: they deliberate token by token, literally writing their way to an answer. A System One Model is the opposite shape — it answers the way a trained reflex fires, but a reflex whose confidence has been trained to mean something. The term was introduced by TypeSafe AI in the Jev launch post observed. It is currently the only name the category has, which is why this independent site uses it analysis.
input: one piece of unstructured text (the state)
+ a list of options YOU define
output: one calibrated probability per option
(single pass, no generated text)
Three properties do the heavy lifting:
1 · No string generation. The model never writes text, so it cannot invent facts, enum values, or JSON fields that were never in your option list. TypeSafe presents this as the design trade that buys both speed and the no-hallucination property vendor claim.
2 · Calibrated probabilities. Jev is trained with Reinforcement Learning for Calibrated Decisions (RLCD), meaning the confidence numbers are meant to be usable as numbers — a 0.8 is intended to be right roughly eight times in ten vendor claim. Calibration is what lets downstream software set thresholds and act automatically.
3 · Single pass. One forward pass per decision instead of autoregressive, token-by-token generation. This is the mechanical source of the claimed ~100× speed and efficiency gap on decision tasks vendor claim.
| LLM (System 2) | System One Model | |
|---|---|---|
| Output | Generated text — prose, JSON, code | Probability distribution over options you defined |
| Process | Autoregressive, token by token | Single pass vendor claim |
| Can invent facts? | Yes — hallucination is intrinsic to text generation analysis | No text generated, so TypeSafe claims no vendor claim |
| Confidence numbers | Token likelihoods — not calibrated for being right analysis | Calibrated by training (RLCD) vendor claim |
| Latency profile | Grows with output length; the calculator on this site models ~700–900 ms per decision call as typical defaults analysis | Claimed ~100× faster than LLMs on decision tasks vendor claim; real deployed latency TBD |
| Best at | Open-ended reasoning, writing, multi-step tool use analysis | High-volume closed-set decisions: routing, moderation, scoring, triage analysis |
Read as: not a replacement — a new slot in the stack. Keep System 2 for reasoning and drafting; put System One in the hot loop where every millisecond and every dollar per thousand decisions counts. analysis
Not a chatbot. It cannot answer free-form questions or write a sentence — text generation is the thing it gave up vendor claim.
Not an embedding model. Embeddings return anonymous vectors for similarity search; a System One Model returns named probabilities over options you chose, ready to act on analysis.
Not a classifier you hand-trained. It is the closest classical cousin, and worth being honest about. Differences: you specify options in natural language at call time instead of collecting a labeled dataset per task, and calibration is built into the vendor's training (RLCD) rather than something you bolt on afterward analysis (I/O shape per vendor claim).
Not independently benchmarked. No public independent replication of the speed or accuracy claims existed as of 2026-09-20 observed. This site will add measured numbers when they appear.
| Item | Status |
|---|---|
| First commercial model | Jev, by TypeSafe AI — announced Sep 15, 2026 observed |
| Access | Waitlist/early access at TypeSafe; listed on Cloudflare Workers AI as typesafe/jev observed |
| Public pricing | None published as of 2026-09-20 observed — model your own assumptions in the calculator |
| Independent benchmarks | None public as of 2026-09-20 observed |
| Open-source lookalikes | Toy-scale research starters only (e.g. jevlike, MIT) — not Jev-equivalent, by their own READMEs observed |
The head-to-head with LLM tool use: mechanism, failure modes, cost shape.
Structured output compared: what a format guarantee is worth vs a decision guarantee.
Model your decision workload against a System One approach.
Pick a scenario, define options, get the schema + call snippets.