TypeSafe AI’s Jev landed in the middle of one of the most crowded weeks of AI releases this year, and it does the opposite of everything else that shipped this month. Anthropic, Google, and OpenAI all put out new models built to reason better and hold longer conversations. Jev is built to never hold a conversation at all.

Instead of generating text, it takes a set of questions with defined answer options, similar to writing a multiple choice quiz, and returns a decision with a probability attached. No paragraphs, no back and forth.

Access has been invite only. I applied when the news broke and got in a few hours later, before I’d had a chance to run anything through it.

What TypeSafe actually built

TypeSafe AI came out of two years in stealth on September 15 with $40 million in seed funding led by DCVC. The company was founded by Diogo Almeida, a former OpenAI researcher who worked on RLHF, InstructGPT, ChatGPT, and GPT4, along with cofounders Erik Gafni and Sasha Sheng. Forbes reported the round values the company at $200 million.

Jev is what TypeSafe calls a “System One Model,” a name borrowed from Daniel Kahneman’s split between fast, intuitive System 1 thinking and slow, deliberate System 2 reasoning. It’s trained with a method the company calls Reinforcement Learning for Calibrated Decisions, aimed at getting the model’s confidence scores to actually match how often it’s right, rather than just sounding confident.

Every question you send Jev has to take one of three shapes. Choice picks an option from a list you define. Score rates something against a rubric you set. Noul answers a yes or no question and returns it as a probability between 0 and 1. You can bundle a lot of these questions into one call, and TypeSafe says adding more barely changes the response time since each one is evaluated separately.

Where the limits are

This is where the onboarding screen I saw after logging in actually lines up with what independent reviewers are saying too. Jev is explicitly weaker than large reasoning models at tasks that need extended reasoning, like math or chess. It has no deep knowledge of specialized domains unless you feed it context yourself. And it can’t chat. To get an answer out of it, you have to define the shape of that answer in advance.

That’s a real tradeoff, not a hedge. Reviewers who’ve tested it describe Jev as a decision engine rather than a smaller ChatGPT: good for classification, ranking, or routing calls a piece of software needs to make, not for generation, explanation, or open-ended planning.

The numbers need a caveat

TypeSafe is pricing Jev at $0.042 per million input tokens with output tokens free, and claiming 20 to 200 times faster and 40 to 400 times cheaper than comparable LLMs. Those figures are worth repeating because they’re dramatic, but they’re also entirely vendor published. There’s no independent benchmark yet, and TypeSafe hasn’t named a single customer using Jev in production. Early write ups have shown promising demos, including one writer at Every who ran 21 questions against 37 documents in under a second, but a fast demo isn’t the same as a verified claim at scale.

Why this matters if you build software

The pitch that seems to be landing with developers isn’t about intelligence, it’s about fit. A lot of software already makes small judgment calls that get bolted onto a large language model today: is this support ticket urgent, does this input violate a policy, which category does this record belong in. Those are narrow decisions being asked of a model built for open ended conversation. TypeSafe’s bet is that splitting that work out to a purpose built model is faster, cheaper, and more consistent, with a confidence score attached so your code knows when to act automatically and when to hand the decision to a human.

What’s next for me

I’ve got access now, but I haven’t run a real test through it yet. I’ve already got a shortlist of small, bounded decisions in a couple of my own projects that look like a good fit for exactly the kind of thing Jev is built for. Once I’ve actually put it through a real test, I’ll share what happened, including where it held up and where it didn’t.

FAQ

What is a “System One” AI model?
It’s TypeSafe’s term for a new category of AI model built to make fast, structured decisions rather than generate conversational text. The name references Daniel Kahneman’s distinction between fast, intuitive thinking and slower, deliberate reasoning.

How is Jev different from ChatGPT or other LLMs?
Jev doesn’t generate text at all. You send it structured questions with defined answer types, choice, score, or yes/no, and it returns a typed answer with a probability and confidence score instead of a written response.

Can regular developers get access to Jev right now?
Access is currently invite only through a waitlist on TypeSafe’s site. TypeSafe says it’s bringing developers off the waitlist as quickly as it can.