Jev, explained independently
What is Jev?
Jev is a “System One” model published by TypeSafe AI in September 2026. It does not chat. Given a narrow question, it returns a typed decision — which element to click, which action to take, whether a request matches a rule — with a stated confidence, so that regular code can act on it. This page explains what Jev is, what it has actually been shown doing in public, and what the evidence does not show.
Not that Jev: the other JEV
A different “JEV” is the standard abbreviation for Japanese encephalitis virus, so medical and biology papers dominate some search results for phrases like “Jev model”. That is a naming collision, not this topic. Everything on this site is about the AI model named Jev that TypeSafe AI announced in September 2026 and offers in early access at typesafe.ai.
What Jev actually does
TypeSafe positions Jev against chat LLMs: rather than producing open-ended text, Jev makes bounded judgments — a choice among known options or a classification of known inputs — and returns them in a typed format a program can branch on. The pattern in every public example we have indexed is the same split of responsibilities:
- Jev owns one bounded judgment: pick the next action, rate the candidate, route the request, accept or reject the match.
- Code owns everything else: the loop, permissions, thresholds, validation, retries, and side effects.
TypeSafe's homepage reports headline numbers for this approach — 193.6x faster, 444.6x cheaper, “zero hallucinations”, and early-access pricing of $42 per billion input tokens. Those are the vendor's reported claims, not independently reproduced measurements.
What Jev can do: 29 public records
Our capability directory indexes every public demonstration we could verify, each labeled with an evidence level and whether the record is official (published by TypeSafe itself, currently 3 of 29) or community (published by a third party). Examples of what Jev has been shown deciding:
- Search Google Flights in seconds — E2 evidence, community record.
- Drive macOS from structured screen state — E2 evidence, community record.
- Choose drone tactics from camera-derived state — E2 evidence, community record.
- Complete a StarCraft shareware mission — E2 evidence, community record.
The full list, with sources and stated limits, is in the directory of 29 Jev capabilities.
What the public evidence does not show
This directory exists because most claims about new models are reported, not verified. Three limits apply to everything above:
- A public repository proves an implementation exists. It does not independently verify every outcome its README claims.
- Metrics stay marked Reported unless an editor has reproduced them under recorded conditions.
- Model confidence is not correctness. A typed answer with 0.99 confidence is still a claim.
How we grade evidence is documented on the editorial method page, and every record links to its public sources.
Jev questions, answered
- What is Jev?
- Jev is a "System One" model published by TypeSafe AI. Instead of generating chat replies, it returns bounded, typed decisions — for example which UI element an agent should act on next — with a stated confidence, so deterministic code can act on them.
- Can you chat with Jev?
- No. Jev is not a chatbot. TypeSafe's own positioning is that Jev does not hold conversations; it answers narrow decision questions inside a program, while the surrounding code keeps control of permissions, thresholds, validation, and side effects.
- Who makes Jev?
- TypeSafe AI, the company behind typesafe.ai. This page and the typesafeai.app directory are independent, unofficial documentation of what Jev has publicly done — not affiliated with, or endorsed by, TypeSafe.
- Is "Jev model" the same thing as the virus JEV?
- No. JEV is also the standard abbreviation for Japanese encephalitis virus, which is why medical papers dominate some search results for "jev model". The Jev discussed here is the TypeSafe AI model announced in September 2026.
- What can Jev do?
- The typesafeai.app directory tracks 29 public records of Jev making decisions in real projects — from driving a macOS agent from screen state to gating risky coding-agent tool calls — each with a source, an evidence level, and stated limits.
- How much does Jev cost?
- TypeSafe's homepage lists early-access pricing of $42 per billion input tokens and markets Jev as dramatically faster and cheaper than chat LLMs for workflow tasks. These are the vendor's reported numbers; this directory has not independently reproduced them.