What is Jev
Jev is the first System One model built by TypeSafe AI. The official definition of System One models: fast, structured decisions for software — you provide a state and typed questions, and get back structured, machine-consumable answers.
One-line distinction: LLMs generate text; Jev returns decisions.
Official positioning
Section titled “Official positioning”Per the official docs (verified 2026-09-22), Jev’s contract is:
- Input:
state(a string, object, or array — the context) +questions(a set of typed questions) - Output: a typed answer per question — the selected option, a score, probability distributions, confidence
- One endpoint:
POST https://api.typesafe.ai/v1/systemone
It does not generate natural-language text. You cannot ask it to write, translate, or summarize; it only answers within the three question types (see Capabilities).
TypeSafe AI was founded by Diogo Almeida, previously at OpenAI working on ChatGPT-related systems (media coverage; also self-described in the launch blog). The reported $40M seed round is a media figure this site has not independently verified.
Where the names come from
Section titled “Where the names come from”The launch blog explains both (verified 2026-09-22):
- Jev ← the Jevons paradox: making something efficient increases its total consumption — the official metaphor for “once decisions become extremely cheap, software will start deciding in places it never dared to before”
- System One ← “System 1” from Kahneman’s Thinking, Fast and Slow: fast, intuitive, narrow judgment — complementary to “System 2” (slow reasoning, i.e. LLMs)
Why it’s fast and cheap
Section titled “Why it’s fast and cheap”Official blog claims: 70–500 ms typical latency, free output tokens, $0.042 per million input tokens. Reasons (official framing):
- Narrow output space: answers are enums and probabilities, not free text — decoding is trivial
- Parallel sampler: the training/inference stack is optimized for huge numbers of small judgments in parallel
- RLCD training: Reinforcement Learning for Calibrated Decisions, the official training method that teaches the model to return decisions with calibrated probabilities instead of generated text
What it is not
Section titled “What it is not”- Not a chat model: no dialogue, no role-play, no long-form generation
- Not open-weight Jev: what’s open is the API and official SDKs; SemIf, kev, and NanoJev are community replicas (see Ecosystem)
- Not “never wrong”: the official “can’t hallucinate” means answers are constrained to the question’s type (no fabricated text outside the format) — not that judgments are always correct (see Limitations)
Next steps
Section titled “Next steps”- Capabilities: the three question types and parallel questions
- Get started: sign up, get a key, send your first request
- FAQ: pricing, free credit, context limits, and more