Jev Guide
A new AI model dropped on September 15, 2026, and it made noise by refusing to make any. Jev — the first model from TypeSafe AI — doesn't chat, doesn't draft, doesn't generate a single sentence of prose. You hand it a state and typed questions; it hands back structured answers with calibrated probabilities. That's the whole interface, and for automation work, it turns out to be the right one.
This guide is the front door to everything we cover on Jev: what the model is, where the idea comes from, how the API works, what the performance claims really mean, and where to go next.
What Is Jev AI?
Jev is the first of a model class TypeSafe AI calls System One models — named for the fast, intuitive mode of thinking in Kahneman's Thinking, Fast and Slow, as opposed to the slow, deliberate System 2 reasoning that LLMs behave like. Where a large language model answers in generated text, Jev answers in typed values: an option picked from your list (Choice), a rating on your scale (Score), or a probability between 0 and 1 (Noul). Every answer carries confidence scores, and responses outside your schema are mathematically impossible.
The company behind it was founded in 2024 by Diogo Almeida, who spent about four years at OpenAI building the RLHF research behind InstructGPT, ChatGPT, and GPT-4. His verdict on chat models — "we have lightning in a bottle, and yet it is not useful" — is the origin story of the product. The launch came with a $40 million seed round led by DCVC.
For the full story — the Kahneman connection, the Jevons paradox behind the name, the three question types with runnable code, and an honest read of the speed claims — start with What Is Jev?, the complete guide.
Jev at a Glance
| Fact | Value |
|---|---|
| Developer | TypeSafe AI (San Francisco, founded 2024) |
| Released | September 15, 2026 (early access) |
| Model class | System One model — decision-oriented, non-text-generating |
| Question types | Choice, Score, Noul |
| Response time | 70–500 milliseconds end to end |
| Input pricing | $0.042 per million tokens |
| Output pricing | Free |
| Training method | RLCD (Reinforcement Learning for Calibrated Decisions) |
| Also available on | Cloudflare Workers AI (typesafe/jev) |
Where to Find Jev
| Entry | Address | What it's for |
|---|---|---|
| Official site | https://typesafe.ai/ | Company hub, launch announcements |
| Documentation | https://docs.typesafe.ai/ | Primitives, patterns, SDKs (Python & JavaScript) |
| Early access | waitlist on the official site | Direct API access, opened gradually |
| Cloudflare Workers AI | typesafe/jev | Run Jev without waiting for the waitlist |
| Wikipedia | Jev (AI model) | Neutral overview with sources |
One caution for searchers: the all-caps acronym JEV belongs to the Japanese encephalitis virus and has for decades — health-authority pages dominate results for "what is jev" without the "ai". If that's what you were looking for, the CDC is the right door. The model is always worth an extra keyword.
What This Guide Covers
The collection is young and grows with the product. Right now:
- What Is Jev? — the complete explainer: origins, the three primitives with runnable examples, Jev vs LLM trade-offs, performance claims with their fine print, and when to reach for each.
On the way as early access widens: a hands-on guide to getting access (waitlist, Cloudflare, playground options), a developer walkthrough of the GitHub ecosystem and SDKs, and a close comparison of Jev against traditional LLMs for real workloads. For the broader model landscape that Jev sits inside, see our AI guides; for the open-weight alternative that TypeSafe's own eval notes keep name-checking, the DeepSeek guide is here too.
FAQ
Is Jev an LLM?No — it's transformer-based and reads natural language, but it never generates text. Outputs are typed values defined by your question schema, which is why TypeSafe files it under a new class, System One models.
How fast is Jev?End-to-end responses run 70–500 milliseconds. TypeSafe reports 40–200x faster and 40–400x cheaper than frontier LLMs on comparable tasks, with peaks of 193.6x/444.6x on its own workflow evals — figures the company itself calls the high end of real-world gains.
Can I use Jev for free?Output tokens are free on the direct API; input costs $0.042 per million tokens. Cloudflare Workers AI lists Jev with its own billing, which is often the fastest way to run a first call.
Who made Jev?TypeSafe AI — founded by Diogo Almeida (ex-OpenAI, RLHF/ChatGPT research), Erik Gafni, and Sasha Sheng, with a $40M seed round led by DCVC.
Whether Jev turns out to be a category of one or the first of many, the idea it stands for — AI that outputs decisions instead of prose — is worth understanding now. Start with the explainer, run the three-primitive example, and you'll have the whole picture inside an hour.