Three different people sent me the same link this week, which is usually a decent signal that something is either genuinely interesting or being very well marketed. In this case I think it’s a bit of both.
The thing is called Jev, from a startup called TypeSafe AI. It went generally available on Sunday. In its first 24 hours on Vercel’s AI Gateway it was picked up by about 13% of their paying teams – faster than any model they’ve ever carried. Worth ten minutes of your time to understand what it is, and about the same again to understand what it isn’t. If you’re looking to get your hands dirty – you’re out of luck – as of Tuesday 22nd – TypeSafe AI isn’t taking on any more tire kickers.
What it actually does
Every AI model most of us have used in the last four years is, fundamentally, a writer. You ask it something, it writes you something back. That’s the interaction model and we’ve stopped noticing it.
Jev refuses to write anything at all. You hand it some raw material – an email, a support ticket, a log entry, a document – and you ask it a question where you’ve already listed the possible answers. Is this billing, technical or sales ? On a scale of routine to urgent, how bad is this ? Is this person angry, yes or no ?
It picks one. That’s the whole output. No explanation, no “certainly, I’d be happy to help”, no prose. Just the answer, plus a number telling you how confident it is.
“Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out.”
Diogo Almeida, TypeSafe AI
It’s the difference between asking a colleague to write you a report and asking them to just point at the right folder. And if you’ve ever actually looked at what software spends its time doing – pointing at the right folder is most of the job.
Why anyone is excited
Two reasons and they’re both unglamorous.
It’s fast. Answers come back in well under half a second. A chat model asked the same question takes seconds, because it’s busy composing sentences one word at a time before it gets round to the bit you wanted. Jev produces the whole answer in one go. Ask it five questions about the same document and it answers all five in roughly the time it takes to answer one – which turns out to matter a lot, and I’ll come back to that.
It’s cheap. Roughly four cents per million words in, and the output is free. That’s not a discount, it’s a different order of magnitude.
Put those together and a whole category of thing that was previously too expensive to bother with becomes trivial. Every decision point buried inside a piece of software – which queue does this go in, does a human need to see this, should we escalate, does this look dodgy – can now have a bit of judgement applied to it for essentially nothing. The model is named after William Stanley Jevons, the Victorian economist who noticed that when something gets cheaper we don’t save money, we just use far more of it (Jevon’s Paradox). They are not being subtle about the thesis.
The bit where I stop being impressed
TypeSafe’s homepage says Jev is 193 times faster and 444 times cheaper than the alternatives. Those are lab numbers – a single isolated decision, measured against a single call to a big expensive model, on test cases the company designed itself. To their credit they more or less admit this in their own materials.
The real world figures published so far are a lot more modest. A production document pipeline saw about six times faster. An independent test came in around five. And in the most honest datapoint of the lot, one of TypeSafe’s own employees swapped a single step of a live pipeline over to Jev and measured a 15.9% speed improvement and a 30% cost reduction.
Not 193 times. Fifteen point nine percent – because replacing one link in a chain doesn’t make the chain much shorter. Anyone who has ever done performance optimization knows this feeling.
For the record – 15.9% is a good result, and the honest range of three to thirty times faster is an excellent one. I’d have more confidence in the product if the marketing had more confidence in the actual numbers.
The other claim doing heavy lifting is that Jev “can’t hallucinate”. This is technically true and practically misleading. Because the answers are listed up front, it can’t invent a fourth option when you gave it three. What it absolutely can do is pick the wrong one of the three and tell you it’s 90% sure. That isn’t hallucination in the technical sense but it’ll ruin your afternoon in exactly the same way.
And that confidence score – the thing that’s supposed to make this trustworthy – is the part I’d test hardest. Someone ran an independent check on it. On standard public test sets, the confidence numbers were honest and well behaved. On a set of invented support tickets the model couldn’t possibly have seen before, they drifted badly, and in different directions depending on the type of question – overconfident when picking from a list, underconfident on yes/no. Useful, but you’d want to calibrate it against your own data before you build anything load bearing on top of it.
What I’d actually do with it
Assuming you’re building something, a few things worth knowing :
Break the hard question into several small ones. This is the genuinely interesting finding. On a single difficult judgement call, a decent small chat model beats Jev comfortably. Decompose the same judgement into five simple typed questions and Jev wins – it went from 62% to 95% on one phishing test while the chat model it was up against actually got slightly worse. That’s an unusual property and it changes how you’d design around it.
Don’t trust the confidence number out of the box. Test it on your data, per question type.
It’s a component, not a system. It replaces the decision, not the pipeline. See the 15.9%.
Price your own workload. One developer found it 10-20x more expensive than the alternatives for his particular shape of traffic. TypeSafe also say straight out that they can’t prove the price isn’t subsidised, which I appreciate them saying and would still plan around.
Today versus where this goes
Today it is good at sitting inside software and making small, fast, cheap, repetitive decisions that previously needed either a rigid rule or an expensive model. Sorting, routing, triage, flagging, screening, guarding an agent before it does something silly. It is not good at anything requiring explanation, nuance, or a question you didn’t anticipate. It can’t tell you why. And it runs only on TypeSafe’s servers – no published research, no model you can run yourself, no option to put it inside your own building. In my day job that last one is the end of the conversation, not the start of it, and I suspect that’s true for a lot of regulated industries.
Where this goes is the more interesting question. I don’t think the bet is that Jev gets smarter. The bet is that “AI model that returns a decision instead of an essay” becomes a standard shape that everybody offers – it doesn’t look especially hard for the big labs to copy, and open source versions are apparently already turning up within days of launch. If that happens, the lasting contribution here isn’t this particular model, it’s the idea.
This is directionally very encouraging – we’re burning tokens at an alarming rate and using a warehouse sized brain to answer trivial classification questions – remember early ChatBots trying to auto-complete math questions and failing – now they just reach for a Python CLI. Horses for courses.
Jev went GA on 21 September 2026. Everything above was accurate the day after, which in this field gives it a shelf life of about a fortnight.
