Jev is TypeSafe AI's System One model: it returns typed, calibrated decisions instead of text, in 70-500ms, for routing, guardrails, and classification.
What Is Jev?
Most AI models — ChatGPT, Claude, and the rest — work by writing an answer one word at a time, like a very fast typist. Jev doesn't type. Ask it a yes/no question and it hands back a probability; ask it to pick from a list and it hands back the pick plus how sure it is. No sentences, no explanation, just an answer your code can act on immediately — which is also why it's roughly 100x faster and cheaper than asking a chat model the same question.
A real request looks like this: you send Jev a situation — say, a support message that reads "Help! My payouts have been failing for 3 days" — plus two questions: is this urgent? and which team should handle it? Jev answers both at once: 97% confident it's urgent, and "technical" with an 83% probability over "billing" and "sales". That's the whole interaction. There are three question shapes: Noul (yes/no, as a probability), Choice (pick one from a fixed list, with the odds on each option), and Score (grade against a rubric you define). Because every answer is locked to a schema you declared upfront, Jev can't reply with a made-up option or broken JSON — though it can still be confidently wrong, so the standard pattern is to act automatically above a confidence threshold and fall back to a human (or an LLM) below it.
How It's Priced and Accessed
Input tokens cost $0.042 per million; output tokens are free. Jev is in early access: sign up for the waitlist at typesafe.ai, then create an API key in the console once admitted. Official SDKs cover Python 3.10+ and Node 20+, with community clients for Rust, Ruby, Go, and Elixir, plus gateway model IDs on Vercel AI Gateway and Cloudflare Workers AI. TypeSafe AI's own benchmark claims roughly 190x faster and 440x cheaper than a comparable LLM on these decision workloads — that's their number on their workflow, not an independently verified figure, and early coverage has flagged it as a ceiling case rather than a typical result.
What Developers Are Building With It
- LangChain shipped a
TypeSafeClassifierintegration so a Jev call can sit inside a LangChain agent loop for routing and classification instead of a chat model. - Browserbase is using Jev to power browser-use agent decisions at a fraction of the per-call cost of an LLM.
- Independent builders have wired Jev into a live trading agent and an email-triage pipeline running at scale.
- A community tool called save-token-jev uses Jev's Noul primitive to decide, per past tool call, whether it's still relevant to a coding agent's context window — trimming context without an LLM summarization call.
Who's Behind It
TypeSafe AI is led by three named co-founders: CEO Diogo Almeida, who co-invented RLHF and InstructGPT (the methods behind ChatGPT and GPT-4) and previously worked at Google Brain; COO Sasha Sheng, an ex-Meta/FAIR research engineer who worked on News Feed and AI Research; and CTO Erik Gafni, a repeat founder (Ravel, multi-modal AI for DNA sequencing) and early employee at Invitae and Freenome. The wider team is drawn from OpenAI, Google Brain, Meta/FAIR, Stripe, Airbnb, Plaid, and Docker, working in person in San Francisco.
Where It Fits
Jev isn't a drop-in replacement for a chat model — it doesn't generate text. Teams are pairing it with an LLM: Jev handles the fast, repeated judgment calls (route this ticket, is this prompt risky, score this candidate), while the LLM still handles open-ended generation and reasoning. Typical use cases are support/prompt routing, guardrails and injection detection, agent tool-selection and context management, real-time decisions inside game or robotics control loops, search reranking, and batch classification where free output tokens make high-volume tagging cheap. The short version: anywhere your code currently burns a full LLM call just to get a yes/no or a pick-one answer, that's a Jev-shaped hole.