TypeSafe AI Launches Jev, a New AI Model That Makes Decisions Instead of Writing Text

A San Francisco startup called TypeSafe AI has introduced Jev, a new kind of AI model that does not generate text or images at all. Instead, it takes a question and a description of a situation and returns a typed, structured answer along with a confidence score. Developers describe it as a "smart if statement" or a classifier with brains, built to sit inside software rather than to chat with people.

Jev came out of stealth on September 15, 2026, alongside a $40 million seed round led by DCVC. TypeSafe was founded in 2024 by Diogo Almeida, a former OpenAI researcher who has been credited as one of the people behind ChatGPT. The company says Jev spent roughly two years in development.

TypeSafe describes Jev as a "System One" model, a term borrowed from Daniel Kahneman's Thinking, Fast and Slow. It is meant to be quick and intuitive, in contrast to slower "System Two" reasoning models. According to reports, it is a transformer-based model trained entirely on synthetic data using a technique Almeida calls reinforcement learning from calibrated decisions.

Instead of paragraphs, Jev returns outputs such as:

TypeSafe says Jev can respond in as little as 150 milliseconds, and its own benchmarks claim large speed and cost advantages over general-purpose chatbots on decision tasks. The model is priced at about $0.042 per million input tokens. Developers can use it through TypeSafe's API (the current version is jev-1.13.0) and through platforms including Vercel AI Gateway, Cloudflare and OpenRouter.

These figures come from the company and early testers, and independent testing is still emerging.

Early adopters have used Jev for routing, classification, safety checks and game-playing experiments. Reported examples include:

A community site called Jevable has sprung up to collect prototype apps.

The launch drew attention across the tech community. Almeida's launch post on X reportedly passed 38 million views within a week. AI researcher Andrej Karpathy described Jev as exposing demand for fast, low-latency models of modest intelligence that had been overlooked during the race for ever-smarter systems. Others were more sceptical, with one widely shared reaction calling it just a very smart switch statement.

Jev cannot explain itself in words, hold a conversation or act as a perception layer for raw sensor data, and it is not intended for real-time control loops. Its creators and testers say it works best when the situation is first turned into a clear description that the model can judge.

If Jev's approach catches on, it points to a new tier in the AI stack: small, cheap decision models handling routine judgement calls, with large language models reserved for harder reasoning and conversation. Commentators expect competitors to follow now that the use case is proven.

Details are drawn from media reports and company claims, and benchmark figures have not all been independently verified.