Most AI models today write text. You ask a question, and they reply word by word. That works great for chat. But it is slow and messy when your code just needs a quick decision.
Jev takes a different path. It is a new AI model from TypeSafe AI, and it does not write text at all. It gives your software direct answers with a confidence score attached.
This guide explains what Jev is, how it works, where it shines, and where it does not.
What is Jev?
Jev is the first model from TypeSafe AI, a San Francisco startup founded in 2024. It launched in early access on September 15, 2026, along with a $40 million seed round led by DCVC.
The company was co-founded by Diogo Almeida, who worked at OpenAI on RLHF, InstructGPT, and ChatGPT before leaving in 2024.
TypeSafe calls Jev a System One model. The name comes from Daniel Kahneman's book Thinking, Fast and Slow. System 1 is the fast, instinctive part of thinking. System 2 is the slow, careful part. Jev is built for the fast kind of decisions.
The name "Jev" comes from William Stanley Jevons, an economist. His idea was simple: when something gets cheaper, people use much more of it. TypeSafe expects the same to happen with AI decisions.
How Jev works
Here is the basic flow:
- You send a state. This is the situation, like a customer message or game data.
- You ask typed questions. For example, "Which team should handle this?" with a fixed list of options.
- Jev returns answers with probabilities. Every answer comes with a confidence score.
Jev supports three kinds of questions:
- Yes or no: Is this message spam?
- Choice: Which category fits best?
- Score: How urgent is this, from low to high?
Here is a simple example. A customer writes: "I upgraded yesterday but now I can't access the features I paid for."
Jev might return:
| Category | Probability |
|---|---|
| Technical | 64% |
| Sales | 23% |
| Billing | 13% |
| Cancellation | 0% |
Your code picks "Technical." But it also sees the message is a bit unclear. So you can decide: act on it, or send it to a human.
Why "type-safe" matters
With a normal LLM, you ask for JSON and hope it comes back correct. Sometimes it adds extra text. Sometimes it makes up a field. Your code then breaks.
Jev cannot do that. You define the possible answers in advance, and Jev can only pick from them. There are no type errors and no made-up options.
This is where the company name comes from. The output is always the shape your code expects.
Speed and cost
This is the headline claim, so it is worth being careful.
- Speed: TypeSafe says Jev responds in 70 to 500 milliseconds. Frontier LLMs can take seconds or more.
- Cost: Input costs $0.042 per million tokens. Output is free.
- Headline numbers: TypeSafe reports up to about 194x faster and 445x cheaper than LLMs on its own workflow tests.
The company is open that these numbers are likely at the high end of real results. Its own team built the test workflows, and it admits some bias is possible. So treat the big numbers as a best case, not a promise.
Where Jev is a good fit
Jev works best when your code needs a fast, clear decision. Good examples include:
- Smart routing: sending support tickets to the right team.
- Classification at scale: tagging millions of records cheaply.
- Scoring and checks: rating content, spotting spam, or checking another AI's output.
- Real-time apps: games, bots, and simulations where every millisecond counts. TypeSafe has shown demos like a Doom bot and a Wikipedia racing game.
- Confident automation: act when confidence is high, and hand off to a human when it is low.
Where Jev is not a good fit
Jev is not a replacement for ChatGPT or Claude. Keep these limits in mind:
- It does not write text. No emails, no summaries, no code.
- It has no memory. You send the full state with every request.
- Choice lists have a limit. A single choice question supports up to 255 options.
- It is closed and hosted. The model runs on TypeSafe's servers. You call it through an API.
- It is still early. Most results come from TypeSafe's own tests. Independent reviews are only starting.
Jev vs a normal LLM
| Normal LLM | Jev | |
|---|---|---|
| Output | Free text | Typed answers with probabilities |
| Speed | Seconds | 70 to 500 ms |
| Output format errors | Possible | Not possible |
| Confidence | Often unreliable | Built into every answer |
| Best for | Chat, writing, coding | Fast decisions inside software |
The smart approach is to use both. Let an LLM handle open-ended work. Let Jev handle the many small, repeated decisions around it.
How to try Jev
Jev is in early access. You can sign up on the TypeSafe AI website. It is also available through gateways like Cloudflare Workers AI, and there is a LangChain integration for agent builders.
Final thoughts
Jev is not a smarter chatbot. It is a different tool for a different job. If your software makes lots of small decisions, like routing, tagging, or scoring, Jev could make them faster, cheaper, and more reliable.
Just keep the hype in check. The early numbers look strong, but they mostly come from TypeSafe itself. Test it on your own data before you depend on it.
FAQ
Is Jev an LLM?
Not in the usual sense. It does not generate text. It returns structured answers with probabilities.
Can Jev hallucinate?
It cannot invent answers outside the options you define. It can still pick the wrong option, which is why the confidence score matters.
Is Jev free?
Output is free. Input costs $0.042 per million tokens.
Who made Jev?
TypeSafe AI, co-founded by former OpenAI researcher Diogo Almeida.
Top comments (0)