TYPESAFE'S JEV ANSWERS IN 70 TO 500 MILLISECONDS

Jev, a new model from TypeSafe AI, answers a typed question in 70 to 500 milliseconds and writes nothing to do it. No paragraph, no token by token buildup, just a decision and a confidence score. That is the entire pitch, and it is aimed at the exact moment most AI models handle worst: the second an app needs one clean answer to keep running.

TypeSafe opened early access to Jev on September 15, after two years spent quietly building it. The company calls Jev a System One model, its own term for software that evaluates a set of possible answers in one pass instead of generating text word by word. A support ticket that reads I was charged twice and I need this fixed right now does not need an essay back. It needs three answers: which team, how urgent, does a human need to step in. Jev is built to return exactly that, nothing more.

70 to 500 Milliseconds, Zero Tokens Out

TypeSafe prices Jev at 0.042 dollars per million input tokens and charges nothing for output, since there is no generated text to bill for. Anthropic lists a base input price of 10 dollars per million tokens for Claude Fable 5.1 by comparison, roughly 238 times higher for the same volume of input.

Jev processes a block of program state and a set of typed questions in a single forward pass rather than decoding one token at a time. It does not draft, revise or pad; it evaluates every question in parallel and returns a labeled answer with a probability attached, the way a multiple choice sheet gets filled in rather than an essay graded. TypeSafe reports those 70 to 500 millisecond response times and claims speedups of 40 to 200 times over a standard language model on structured queries, numbers the company has published but that independent benchmarks have not yet confirmed line by line.

Diogo Almeida Spent Two Years Building the Opposite of ChatGPT

Diogo Almeida founded TypeSafe after helping invent RLHF and working on InstructGPT at OpenAI, the same reinforcement learning technique that still guides models like GPT 6 Astra, which cleared all 48 levels of Neal.fun's robot test this month. He spent nearly two years in stealth before Jev's September 15 release, funded by 40 million dollars in seed money led by DCVC.

Almeida has argued publicly that reliability, not raw capability, is the real gap in deploying AI inside software. A chatbot that occasionally refuses to answer or drifts off script can break a production system even when it is smart enough to pass a benchmark. Jev is his answer to that argument, a component built to behave like infrastructure a program calls, not a conversational partner a person types to.

Program State Goes In, a Probability Comes Out

A support ticket, a line of chat, or a stream of sensor data goes in as program state alongside a short list of possible answers, and Jev returns which answer is most likely along with a confidence score attached to it. Developers have pointed it at routing support tickets to the right team, choosing an AI agent's next tool mid task, flagging a risky command before it runs, and filtering documents before they ever reach a larger, pricier model.

That last use case draws the most attention, since it turns Jev into a cheap first pass in front of models actually built to write out an answer, the reasoning kind of work Jev was never designed to touch. The same one pass judgment already runs sneaker authentication tools that call a pair real or fake in under a second, no essay required, just fashion's version of the identical binary logic.

Vercel Adopted It Faster Than Any Model Before

Vercel added Jev to its AI Gateway on September 18 and called it the fastest adopted model in the Gateway's history, three days after launch. Cloudflare, LangChain and Langfuse each shipped integrations in that same window, before most developers outside the AI industry had even heard the name.

The community response arrived just as fast. Within days, three separate awesome jev directories appeared on GitHub cataloging community builds, among them an open source Gmail triage tool that sorts an inbox into needs reply, updates, promos and spam, and an agent that plays Tetris by scoring every possible piece placement as a Jev decision instead of a heuristic, the same kind of tool picking problem SpaceXAI's Grok Bot project is chasing with a different model entirely. None of it proves Jev works everywhere a chatbot does. It proves developers found a narrow enough problem for it fast enough to matter.

Free Output Comes With a Catch

Free output sounds like the whole pitch until you notice what TypeSafe controls underneath it: the taxonomy of allowed answers is set at the API level, so a team that builds its routing logic around Jev's typed responses is writing to TypeSafe's schema, not a portable standard. Switching providers later means rebuilding that answer set from scratch, a lock in cost that a la carte token billing does not carry on its own, even once the per token math favors Jev by roughly 238 times.

That tradeoff is worth paying for the problem Jev actually solves. Try it for a single, well bounded decision node, a support ticket router, a tool picker for an agent, a document filter sitting in front of a bigger model, where 70 to 500 milliseconds and 40 million dollars of runway behind Diogo Almeida's bet on reliability are the whole argument. Skip it for anything that still needs to explain itself in a full sentence; that job still belongs to the model built to write one.