Jev by TypeSafe.ai

Not another chatbot. Fast typed decisions for code, with a confidence number attached.

TypeSafe AI left stealth with Jev, a System One model that answers structured questions in ~100ms instead of writing paragraphs.

By RTTA TeamLast updated 2026-09-18

What it is

Jev by TypeSafe.ai is the model that jumped the developer timeline the week of September 15, 2026. TypeSafe came out of stealth with about $40M seed led by DCVC. Founder Diogo Almeida (ex-OpenAI, RLHF / InstructGPT lineage) shipped Jev as the first public "System One" model: state in, typed probabilistic answers out. No essay. No tool-call poetry. A function call software can branch on. What people are actually reacting to. You send unstructured state plus questions typed as Choice (pick one of N), Score (place on a short scale), or Noul / boolean (yes-no with a probability). Jev answers every question in one parallel pass. TypeSafe quotes end-to-end latency around 70–500ms, input at $0.042 per million tokens, and free output. The company pitch is blunt: it gave up string generation on purpose so the model cannot invent a wrong enum or a chatty refusal when your code needs billing vs technical. Why the name stuck this week. "System One" is Kahneman shorthand for fast judgment. "Jev" nods to Jevons (cheaper intelligence → more demand). Press and YouTube explainers hit the same hooks: Doom bots querying game state ten times a second, support ticket routing, model routers that send easy work to a cheap path and hard work to a frontier LLM, plus side-by-side demos where Jev dumps all probabilities at once while a chat model dribbles tokens. LangChain, Vercel AI Gateway (`typesafe-ai/jev`), Cloudflare Workers AI (`typesafe/jev`), and the official `@typesafe-ai/sdk` showed up within days. Access is waitlisted early access at typesafe.ai, with keys through console.typesafe.ai once you are in. Several writeups say invites landed within a day or two. If you skip the list, the Gateway path is the workaround people mention in demos. Format recipe for the discourse wave: screen-record a tiny classify/route call, put latency and the probability table on screen, then cut to the "this is not ChatGPT" line. Alternate angles are Doom/Wikiracing demos from the launch post, cascade diagrams (Jev first, LLM only when confidence is low), and honest caveats that TypeSafe's speed and cost charts are still self-reported. Not a consumer TikTok dance and not a new Claude or GPT chat model. If the caption says Choice / Score / Noul, System One, or "decisions for software," you are on the right product.

How to try it

  1. Join the waitlist at typesafe.ai or call Jev through Vercel AI Gateway as typesafe-ai/jev. Once you have a key, set TYPESAFE_API_KEY and hit POST https://api.typesafe.ai/v1/systemone (or @typesafe-ai/sdk / experimental_evaluate).

  2. Demo angle: pick one real state string (a support ticket, a refund note, a tool-call proposal). Attach 2–3 Choice / Score / boolean questions with clear criteria. Film the JSON answers and confidence, not a chat reply.

  3. Cascade angle: show Jev routing or gating, then hand only the low-confidence or generative cases to a normal LLM. That is the pattern LangChain and early cookbooks keep repeating.

  4. Search pack: #Jev #TypeSafeAI #SystemOne #AI and the phrase Jev by typesafe.ai. Caution: do not pitch Jev as a writing or coding chat replacement. It does not generate prose.

Jev by TypeSafe.ai — YouTube clip

Walkthrough of what Jev is, how it differs from an LLM, and live classify / router demos.

Open original

Jev by TypeSafe.ai — YouTube clip

Explainer on Jev as a fast, cheap System One model and the tradeoff of no text generation.

Open original

Jev by TypeSafe.ai — YouTube clip

Hands-on try of TypeSafe's System One model Jev after launch.

Open original

Common questions

What is Jev by TypeSafe.ai?
TypeSafe's first public System One model. You send state plus typed questions; it returns constrained answers with probabilities and confidence, usually in under half a second. It does not write chat replies.
When did it launch?
TypeSafe left stealth around September 15, 2026 with Jev in early access and about $40M seed funding led by DCVC. Diogo Almeida is CEO and co-founder.
How do I get access?
Join the waitlist at typesafe.ai, then grab a key from console.typesafe.ai when invited. Some developers skip the list by using Vercel AI Gateway's typesafe-ai/jev model id.
What is Choice, Score, and Noul?
The three question shapes. Choice picks one labeled option (up to 255). Score places the state on a short ordered scale. Noul / boolean returns a yes-no probability. You define the criteria up front.
Should I replace my LLM with Jev?
No. Use Jev for classify, route, score, and gate. Keep a generative model for writing, open-ended reasoning, and anything that needs free-form text.