Insights

Why are we already building with Jev?

TypeSafe AI’s Jev skips the paragraph and hands back the decision, in under half a second, for a few cents per thousand. It’s the motor-function layer AI products have been missing, and we’re putting it to work.

By Innovation Theory

← Go back

In short

Most of what an AI agent does all day isn’t writing. It’s deciding: is this lead real, which queue does this ticket go to, is this task done. Until now, the usual way to get those answers was to pay a large language model to write a paragraph, then pull a yes out of it. Jev, from TypeSafe AI, skips the paragraph. You hand it the facts and a typed question, and it hands back a probability in under half a second, for a few cents per thousand decisions. For a business with AI in its workflows, that’s a faster, cheaper decision layer, and every answer comes back typed and logged where a person can check it. The teams wiring decision models in now will have the faster, cheaper agents next quarter.

TypeSafe AI launched Jev on September 15, 2026, and calls it a “System One” model, a name borrowed from Daniel Kahneman’s fast, intuitive kind of thinking. Jev doesn’t write. You hand it state, like a support ticket, a log line or a form submission, and a typed question: yes or no, pick one of these, or give it a score. It hands back a probability, usually in under half a second, at $0.042 per million input tokens with output free. In practice that’s a few cents per thousand decisions. The company’s founder, Diogo Almeida, helped build the training methods behind ChatGPT, and TypeSafe says it trained Jev with a new method it calls RLCD, reinforcement learning for calibrated decisions.

The launch is refreshingly candid about how it works. TypeSafe’s own evaluations show Jev up to about 200 times faster and 444 times cheaper than frontier models, and independent tests still land at 5 to 25 times, a real step for anything that runs at volume. “Can’t hallucinate” means the answer always fits the format you asked for, which is exactly what software needs from a decision. Like any model, it does its best work when counting, math and dates are handled in code first, and that’s how we build with it. The headline promise is calibration: when Jev says 80%, it should be right about 80% of the time. We wire every build so an owner can watch that hold on their own data.

Here’s why it matters for the businesses we build for. A model that returns the decision directly, fast and cheap, changes the shape of AI work. Decisions become typed, logged and checkable, and the language model is saved for the parts that actually need words. That’s the difference between an assistant that talks about your queue and a queue that sorts itself.

For clients, the win shows up two ways. At thousands of decisions a day, in routing, deduplication or spam triage, the savings are real. For a smaller business, the win is speed and consistency: every lead, ticket and task gets the same question asked the same way, with the answer logged where a person can see it. We run it beside the rules a client already has, keep a fallback in place, and switch over where it wins on their own data.

The demand is already here. Jev landed on Vercel’s AI Gateway on September 16, where roughly 13% of paid teams tried it in the first day, and it’s on OpenRouter and Cloudflare too. TypeSafe paused direct signups on September 22 because of demand. The businesses that get decision models into their workflows early are the ones whose agents will be faster and cheaper while everyone else is still paying for paragraphs.

Source TypeSafe AI: Introducing System One models and Jev

Got a decision your AI could make?
Let’s find out.

Let’s find out
Product
Innovation Score: 64%

The decision layer AI products were missing.

Here’s what a buyer is actually paying for: the inbox that triaged itself before they got to their desk. Jev is the cleanest way we’ve seen to build that. You ask a typed question, you get a number back in under half a second, and it costs cents per thousand. That changes what’s worth automating. Decisions that used to need a person, or a slow and expensive paragraph from a big model, become something you can run on every lead, every ticket, every form. The design that makes it safe to sell is simple: Jev scores, the rules a client already trusts keep routing, and a person sees every place the two disagree. That’s how an owner watches it earn more of the work, on their own data, week by week. We’re already building with it. The businesses that wire this in now get the head start.

“Jev scores, your rules route, and you see every place they disagree.”

Engineering
Innovation Score: 58%

The decision layer, done right. Wire it in.

This is the part we would have drawn ourselves. Most of what an agent does inside a business is decide: real lead or spam, which queue, done or not done. Until now every one of those meant paying a language model to write a paragraph, then digging the yes out of it. Jev hands back the decision itself, typed, with a probability attached, in under half a second, for a few cents per thousand.

Here is how we wire it into a client’s flow. First, every call is logged with its input, its answer and its confidence, so any decision can be traced and replayed. Second, the line for acting is set per decision and tuned on the client’s own records. Third, anything below that line goes to the rules or the person who handles it today, and if the model is ever unreachable, the flow falls back and keeps moving. The language model gets saved for the work that needs words.

“A decision should arrive typed, timed and on the record, and now it can.”

Marketing
Innovation Score: 65%

Your inbox sorted itself while you slept.

Nobody buys a classifier. They buy the morning where the inbox already knows which message matters. The AI reads every incoming request, scores confidence, and puts the ones worth your time at the top. Every decision gets logged, so a person sees every disagreement, and we switch more work over where it wins on your own data. That’s a tool that earns trust by showing its work, not by asking for it.

We’re already building with this. Jev sorts, the rest of the system drafts, and a person signs off. The queue handles itself while you handle the business.

The business down the street is still reading every email themselves. They’re spending their sharpest hour on the dumbest task. You could have a queue that sorts itself and a morning that starts with decisions instead of triage. The only question is how many mornings you want to lose before you call.

“Nobody buys a classifier. They buy the morning where the inbox already sorted itself.”

Design
Innovation Score: 70%

A decision you can see is one you can check.

What we like about Jev is that its answer has a shape. You hand it the facts and a typed question, and it hands back a number. Is this lead real? Something like 0.82. Is this task done? Something like 0.31. Every one of those can sit on a screen where a person reads it, right next to the facts that produced it. That’s the decision layer made visible, and it’s the part of an AI workflow most owners never get to see.

We’re already building with it. The way we show it to a client is a simple log: each decision, the number Jev gave, what the rules would have done, and what actually happened. Over a few weeks that log becomes the client’s own evidence. They watch their 80% answers and see for themselves how often they come true. Nothing rests on trust. A typed decision you can read beats a paragraph you have to interpret.

“A decision with a number on it is a decision a person can check.”

Phil Cowan
Founder
Innovation Score: 100%

What is Jev, the way you’d explain it to a client over coffee?

Jev is an ai for applications, not humans. Having programmed for years, seeing how it works made me more excited than I’ve been in a while. I expect all the top players to follow this lead. Looking forward to local versions to improve as well.

You expect all the top players to follow this lead. Why does it matter for the businesses we build for?

While our brain is thinking right now, there’s a large amount of lower level functions going on to keep things running upstairs. That’s Jev, and the fast typed models that will be coming. Jev is like motor functions.

Would you bet a client build on it today?

yes with a confidence of 100%. I’m already building with it- ai accelerated products will never be the same.

Got an idea it could run?
Let’s find out.

Let’s find out