AILLMsJevSystem Design

Jev: The AI Model That Never Writes a Word

TypeSafe AI's new model does not generate text. It answers typed questions with probabilities in a few hundred milliseconds. Here is what Jev is, how it differs from an LLM, and which of the claims deserve a pinch of salt.

September 22, 2026 · 8 min read

If you have put an LLM inside an app, you know this routine. You have a question with a boring answer, say "should we ship this build?". You send it to the model. The model writes you a small paragraph weighing the risks. Then you write code to dig the word "yes" out of that paragraph, and more code for the day it says "Probably, though I would wait for QA" instead.

We have been paying for essays when we wanted a single value. A new model called Jev is built around that complaint, and it does not write essays at all.

So what is Jev?

Jev comes from TypeSafe AI, a company that came out of stealth on September 15, 2026, with a reported $40 million seed round led by DCVC. The CEO is Diogo Almeida, a former OpenAI researcher who worked on the instruction-following methods behind ChatGPT. His pitch, from the launch post: "Models have been superhuman at chat for years, so where is all the automation?"

Their answer is a new class of model they call a System One model. The name comes from Daniel Kahneman's split between System 1 (fast, intuitive) and System 2 (slow, deliberate) thinking. A reasoning LLM that thinks step by step for ten seconds is System 2. Jev is meant to be the fast gut call. It is also named after the Jevons Paradox, the idea that when something gets cheaper to use, people end up using far more of it.

It is in early access right now. There is a waitlist, and I have only read about it, not used it. Everything below comes from the launch post and the coverage around it.

It only answers three kinds of questions

This is the part that makes Jev feel so different when you first see it. You do not prompt it. You hand it some context and a list of typed questions, and it answers all of them at once. There are three question types:

Noul: a yes or no question. You get back a probability between 0 and 1.
Choice: pick one label from a fixed set. You get the winner plus a probability for every option.
Score: rate something on an ordered scale you describe, like a rubric. You get a number on that scale and a probability for each level.

Picture a listing that just went up on a marketplace app. You ask three questions about it in a single call: is this listing a scam, which category does it belong to, how good is the price on a five point scale. Back come three answers at once: 0.89 for scam, "electronics" for the category, and 2 on the price scale, all in well under half a second. No prompt to write, no listing to read twice.

How it actually differs from an LLM

There are four differences, and they stack.

It does not generate tokens one by one. An LLM is autoregressive. It predicts one token, feeds it back in, predicts the next, and repeats until it is done. Every answer is a long chain of dependent steps. Jev uses what TypeSafe calls parallel, non-autoregressive sampling, so every answer you asked for comes out of a single query instead of a long chain of dependent steps. Ask it three questions about the same context and you still make one call. That is a big part of why it is fast.

It is trained for a different goal. Modern chat models are tuned with RLHF, which optimises for answers humans prefer, and RLVR, which rewards answers that can be verified. Jev uses something TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD). The target is not a pleasant reply. It is a correct decision along with an honest probability. That matters because LLMs are famously overconfident. If a model says 0.9, you want it to be right about nine times out of ten, and that is what "calibrated" means.

Its output is a typed value, not a string. Because the answers come from a schema you defined, a Choice can only be one of your options and a Noul can only be a number between 0 and 1. A malformed answer is not something the model can produce.

It cannot do the things LLMs are famous for. No free-form text, no code, no explanations, no conversation. No image input yet either. It is closer to a very fast classifier that tells you how sure it is than to a chatbot.

People are already trying it on things that look nothing like chat. The launch demos include a bot that picks its next move in a Minecraft-style game, a simulated self-driving car choosing between accelerating, braking and changing lanes, and a drone that decides whether to turn or hover. Every one of those runs in a simulator, so treat them as sketches of the idea and not proof it works on real hardware.

The numbers

TypeSafe claims end-to-end responses in 70 to 500 milliseconds, against multiple seconds for frontier LLMs on the same tasks. Input costs $0.042 per million tokens and output is free, which works out to about $4.20 to run 100,000 requests of a thousand tokens each. Their demos lean on that: 724 ads classified in 40 seconds for about nine cents, 384 news headlines triaged in under 25 seconds for nineteen cents.

The headline claim is 40 to 200 times faster and up to roughly 400 times cheaper than frontier models on these kinds of tasks. Here is where I would slow down.

Those figures come from TypeSafe's own workflow evals, built by their own team. Accuracy was measured as agreement with reference answers from other frontier models, not against human-labelled ground truth. On their own dashboard Jev scores 67.8% against 74.1% for the best comparison model, so it is not winning on accuracy. It is winning on speed and price at roughly comparable quality.

There is some outside evidence. Every ran its own test and found about 25 times faster and 580 times cheaper than Claude on extraction tasks, and called the result "good but not perfect". The speed gain there is smaller than the launch numbers and the cost gain is bigger, but both point the same way. The fair reading is that the direction is confirmed and the exact magnitude is not. Test it on your own workload before believing any multiplier.

"It never hallucinates" needs a footnote

This claim shows up in a lot of headlines, and it is true in a narrow sense. The company reports a 0% type error rate, because Jev can only emit values that fit your schema. It will never return "maybe kind of urgent" for a yes or no question.

But a wrong answer that fits the schema is still a wrong answer. If Jev is 89% sure a listing is a scam and it was actually a genuine seller clearing out old gear, you got a perfectly valid, perfectly wrong result. DataCamp's review makes the same point: the guarantee covers whether the output is well formed, not whether the decision is right. And since Jev explains nothing, you cannot ask why it decided something, which makes debugging and audits harder.

Where I think it fits

I work on LLM reliability, so the calibrated probabilities are the part that interests me most. A confidence you can trust is worth a lot more than a confident-sounding paragraph. It lets you write rules like "act automatically above 0.9, send it to a human or a bigger model below that."

That is also how I would use it, as a layer and not as a replacement. Think about any agent. In Ivo, the personal assistant I am building with LangGraph and MCP, a big share of the steps are tiny decisions: which tool should handle this, is this retrieved chunk actually relevant, does this answer break a rule. Today each one of those costs a full LLM call. A fast decision model could take those steps, and a real LLM would only be woken up for the fuzzy cases that need reasoning or an explanation. Guardrailing another model's output is a natural job for it as well.

It is also worth being clear about what it will not replace. Anything open-ended, anything where the answer set is not known in advance, anything where you need reasoning you can read, stays with LLMs. Requesty's write-up adds that it struggles with arithmetic, counting and date comparisons, and it is not built to resist adversarial input.

Pricing is another open question. Some coverage points out that the current rates might be subsidised, and TypeSafe has no named production customers yet.

Why I find it interesting anyway

For the last few years the main move in AI has been to make one model do everything, and then wrap it in more code to force its answers into a shape. Jev goes the other way. It takes one narrow job, makes the output a value instead of a string, and optimises for speed and honest confidence.

I do not know if TypeSafe's numbers will hold up once people outside the company hammer on it. But the idea underneath it, that a lot of automation needs a quick judgment and not a paragraph, sounds right to me. Even if Jev itself stumbles, I expect we will see more models built this way.

Sources and further reading