•
•
•
•
•
•
•
A beginner's guide to TypeSafe AI's decision model that returns typed answers instead of text
Jev is an artificial intelligence model from TypeSafe AI, a San Francisco company founded in 2024 by former OpenAI researcher Diogo Almeida. Unlike ChatGPT, Claude or Gemini, Jev does not write sentences, answer in prose or generate images. It takes a question and a description of a situation, then returns a typed, structured decision along with a confidence level.
Developers have described it as a "smart if statement" or a classifier with brains. It is not an app you chat with. It is a component you build into software, where fast and reliable judgement calls matter more than eloquence.
| Feature | Details |
|---|---|
| Developer | TypeSafe AI |
| Founder | Diogo Almeida, formerly of OpenAI |
| Announced | September 15, 2026, out of stealth after about two years of development |
| Funding | $40 million seed round led by DCVC |
| Type | Proprietary, transformer-based decision model (not a large language model) |
| Output | Typed answers with confidence scores, no free text |
| Current version | jev-1.13.0 |
| Price | About $0.042 per million input tokens |
| Access | TypeSafe API, Vercel AI Gateway, Cloudflare, OpenRouter |
TypeSafe frames Jev as a "System One" model. The term comes from Daniel Kahneman's book Thinking, Fast and Slow, where System One is fast, instinctive thinking and System Two is slow, deliberate reasoning. Most of today's headline AI models behave more like System Two: they reason step by step and produce long text answers.
Almeida's argument, as reported, is that the AI industry has become extremely good at human language but that language is a poor fit for automation, because software needs structured outputs rather than paragraphs. Jev is built around that gap. It is reportedly trained only on synthetic data using a method Almeida calls reinforcement learning from calibrated decisions, which aims to make the model's stated confidence match how often it is actually right.
A developer gives Jev two things: a question and a state, meaning the context to judge. Jev answers in a fixed format. Commonly reported output types include:
Because Jev does not have to compose a written response, it can reply very quickly. TypeSafe says answers can arrive in as little as 150 milliseconds, which makes it suitable for interactive tools and even games.
Early demonstrations suggest Jev fits best where a rule is too fuzzy for ordinary code but a full chatbot would be slow or expensive:
Vercel reportedly swapped an OpenAI model for Jev in one task and saw results arrive several times faster with better accuracy, though results will vary by use case.
| Jev | Typical LLM chatbot | |
|---|---|---|
| Output | Typed choice, score and confidence | Free-form text |
| Speed | Milliseconds | Seconds or longer |
| Cost | Very low per call | Much higher |
| Best at | Routing, classification, checks | Explaining, writing, reasoning |
| Can explain itself in words? | No | Yes |
TypeSafe's own benchmarks claim large speed and cost advantages over general-purpose models on decision tasks. These numbers come from the company and early testers, and independent evaluations are still emerging.
Jev's launch was widely discussed. Almeida's launch post on X reportedly passed 38 million views within a week. AI researcher Andrej Karpathy said it highlighted demand for a fast, low-latency model of modest intelligence that had been neglected during the race toward ever-higher intelligence. Some developers were more sceptical, joking that Jev is simply a very smart switch statement. A community site called Jevable gathers prototype apps built with the model, and some developers have already started experimenting with open alternatives inspired by it.
Jev points to a possible new layer in the AI stack: small, cheap, specialised decision models handling routine judgement calls, with large language models saved for conversation and hard reasoning. If the approach holds up, it could lower the cost of building AI agents and make automated systems more predictable, since structured answers with confidence scores are easier to test than free text. Commentators expect competitors to follow now that the use case is visible.
Is Jev a large language model?
No. It uses a transformer architecture, but it does not generate text or images.
Can I chat with Jev?
No. It is available through APIs and platforms for developers, not as a chat app.
Who created Jev?
TypeSafe AI, led by Diogo Almeida, who previously worked at OpenAI.
How much does Jev cost?
Reported pricing is about $0.042 per million input tokens.
Is Jev open source?
No. It is proprietary, though community projects inspired by it have appeared.
This article draws on reports and explainers from TechCrunch, The Register, Zapier and Wikipedia, plus launch coverage and community write-ups. Company claims, benchmarks and pricing may change and should be checked against TypeSafe's official materials.
