TL;DR
- Jev cannot write sentences. It picks from answers you supply, charging four and a bit cents per million words read.
- Asking it a question costs about what a database query costs, and takes about as long.
- At two orders of magnitude cheaper, free ad-supported AI apps start to add up. At frontier prices they never did.
Early motor cars were built as horseless carriages. The same high seat, the same buggy springs, and on some models a socket to hold the whip. A whip for the horse that was no longer there. A new technology spends a while wearing the shape of the thing it replaces before it finds its own.
That image opens the manifesto of a company called TypeSafe, and it is aimed at the chat box. AI trained to be a helpful, articulate assistant makes sense if a person is sitting on the other side of it, and produces an AI that needs a human in the loop instead of one that runs quietly in the background. Their line on AGI is blunter: we already have general intelligence, and most software is no more intelligent than it was. Intelligence today is like databases before SQL, powerful but bespoke every single time. The manifesto closes on the line that gives this piece its title: we're building prod, not God.
They shipped it on 15 September. It is called Jev, and it will not write you a sentence.
You ask it a question and it returns an answer with a number attached saying how sure it is. That is the entire range of what it will say, and developers are freaking out about it.
TypeSafe calls the category System One models: a class of AI model built to make fast, structured decisions that software can use directly. You hand one a state, it returns typed answers and probabilities. That is the whole contract.
For two years every lab has improved the same object in the same direction: longer context, deeper reasoning, more modalities, better prose. This one goes sideways, and gets there by taking something away. Give up writing text and you can work out every answer at once instead of one token after another, guarantee the shape of what comes back, and drop cost and latency by two orders of magnitude. We cannot think of another model launched this year whose selling point is something it refuses to do.
How it works
You send a block of text — an email, a support ticket, a page of a contract, the state of a running program — and a question with the answers already listed. Is this command dangerous: yes or no. Which team owns this: platform, billing, or security. How urgent is this, on a scale of one to five. Back comes the answer, plus a number saying how confident it is. You can ask a hundred such questions about the same text at once, and they are all answered together rather than one after another.
A chat model does this too, badly. You ask it to classify something, it writes a sentence explaining its reasoning, wraps the answer in JSON, and you parse the result and hope the shape is right. Most of the time it is. Occasionally it invents a category you never offered, or returns prose where you expected a number, and somewhere three layers down your pipeline something breaks at two in the morning.
A chat model gives you a paragraph and asks you to trust it. This one gives you a number and tells you how much to.
Jev cannot do that, because the shape of the answer is fixed before the model ever runs. The name comes from Kahneman: System 1 is the fast, intuitive mode of thinking, as against the slow deliberate kind. The useful way to hold it in your head is the one TypeSafe uses itself, in a footnote, slightly embarrassed about it: a smart if-statement. Ordinary code cannot decide whether a customer sounds angry. This can, in about a tenth of a second, and then ordinary code decides what to do about it.
What it costs
A frontier chat model charges somewhere between twenty cents and ten dollars per million words of input, and roughly five times that for what it writes back. Jev charges four and a bit cents per million words in, and nothing at all for the answer, because the answer is a number rather than an essay. A round trip takes between 70 and 500 milliseconds instead of somewhere between three seconds and five minutes.
Asking this model a question costs about what a database query costs, and takes about as long.
That is what has developers interested, and the early reports are specific. Vercel's Pranit Sharma told TechCrunch they swapped out an OpenAI model that screens potentially dangerous commands and got answers five to eighteen times faster, more accurately. Mike Taylor at Every ran 777 judgments across 37 documents in under seven tenths of a second, for roughly a quarter of a cent. Armin Ronacher, who wrote Flask, pointed at the use case that probably matters most: deciding, in real time, whether a job is hard enough to deserve an expensive model at all.
Why it is named after a Victorian economist
In 1865 William Stanley Jevons noticed something about coal that everyone else had backwards. Steam engines were getting more efficient. Each one burned less coal to do the same work. The obvious conclusion was that Britain would therefore burn less coal.
Britain burned vastly more. Cheaper steam power meant steam power turned up in places nobody had bothered to put it before, and the new demand swamped the saving many times over.
TypeSafe named the model Jev after him, which tells you how they see this. They are explicit that Jev is not smarter than the big models, and that on the tasks it handles it is roughly comparable. The claim is about cost: two orders of magnitude cheaper and faster at those tasks.
When intelligence cost dollars and seconds, you spent it at the edges of your system. At cents and milliseconds, you can afford to put it in the middle.
Every piece of software you run is full of decisions currently made badly by brittle rules, or made by a person, or simply not made. Which of these five thousand documents should a human look at first. Is this supplier email a complaint or a query. Does this transaction pattern look like the one we saw in March. Those are not questions worth a three-second, three-cent model call each. At a tenth of a second and a fraction of a cent, they might be.
Their own demo is a bot playing Doom by asking the model roughly ten questions a second, at about seven dollars an hour. They cheerfully admit a normal non-AI bot would play better.
Jev is going to change consumer AI
Until now, if you wanted an app that put a model call on every screen — sorting, ranking, spotting, filtering, flagging — the sums did not work. A few video ads and an affiliate link do not cover a model call per user per screen. That left two ways to build: be large enough to swallow the cost, or charge a subscription. Usually both.
Which is why nearly every AI product you have used is from a company with a hundred billion dollars behind it, and nineteen francs a month.
Take two orders of magnitude off the price of a decision and a third option opens up: free, ad-supported, and clever on every screen. Not a general-purpose assistant — something narrower and much more specific. A marketplace that understands what you actually listed. A photo app that files things the way you would have. An inbox that surfaces the four messages that matter this week and the dates buried inside them. Unglamorous, and until now not worth building.
We would expect the interesting ones to be hybrids — an expensive model for the two moments a person really notices, a cheap one for the thousand small judgements underneath. Whether advertising ends up paying for that is a separate question, and we have watched enough ad-funded software slowly get worse to be cautious about cheering too early.
What this means from here
Whether or not this particular company wins, somebody will build this. A great deal of what we currently call "using AI" is a chat model doing a job that never needed prose. It classifies, routes, scores and flags — and we pay it to write a paragraph that we then throw away.
Somebody was going to sell the decision without the essay.
Update: we spent Sunday wiring it into the house
Access came through at the weekend, which is how a perfectly good Sunday disappeared.
Our own assistant — we call it Dbrain; it is a GBrain knowledge store running under OpenClaw on a Synology NAS — has been able to turn the lights on for a while. Hue in some rooms, IKEA bulbs in the others. What it could not do was work out cheaply whether a given sentence was about the lights at all. Every it is too bright in here went to a frontier model and came back when it came back.
Jev now sits in front of all of it and answers exactly that. Is this about the house. Which room. Is it clear enough to act on without asking first. Three questions, one request, an answer before you have finished putting your keys down — and the expensive model only gets woken up when there is genuinely something to think about.
TypeSafe have published their own smart-home demo of the same pattern, and it explains the trick better than we can: fire off every question at once — including the ones you probably will not need — and throw away the answers you do not use. Asking is cheap enough now that guessing wrong costs nothing.
Update 2: we also put it in front of our own search and our tools
Search in our own brain has a reranking step: retrieval hands over a couple of dozen candidates, and something then has to decide which of them actually belongs at the top. That job belonged to Voyage’s rerank-2.5. It is Jev’s now, and the brain itself was never touched.
The obstacle was that GBrain supports exactly one reranker — Voyage, the one we were already paying — and patching GBrain to add another was not available to us: container files get wiped on every recreate, so a patch is a change that un-applies itself the next time anything restarts.
The way through was a piece of luck in someone else's design. GBrain also ships a llama-server-reranker provider, and that provider is not really a product — it is a generic HTTP contract. Post a query and a list of documents, get back indices and relevance scores. Both the model name and the base URL are ordinary config keys.
So we did not patch anything. We wrote a small service that speaks that contract and is answered by Jev underneath, and pointed two config keys at it. GBrain is untouched and does not know the difference.
One pair per call, not all the candidates at once. Batching every candidate into a single state is cheaper and faster, and it is the wrong shape here. TypeSafe publish a list of what the model is bad at, and on it is exactly this: accuracy degrades as the state fills with content irrelevant to the question. Rank twenty-five candidates in one state and twenty-four of them are irrelevant to every question you ask about the twenty-fifth. Doing it properly costs about a tenth of a cent per search, which is the whole argument of this piece arriving as an invoice.
It gates as well as orders. Ranking puts the best passage first. It does not say whether the best one is any good. So every candidate is now also asked, plainly, whether it earns a place in the answer at all, and the ones that do not are dropped rather than padded in behind the ones that do. A model that returns a probability instead of a paragraph is the right shape for that question, and cheap enough to ask of every passage rather than the top few.
It fails open, and that is the opposite of our other use. The classifier that decides whether you are talking about the lights fails closed — if it cannot answer, nothing happens. This one does the reverse: a pair that fails keeps its upstream position, and a gate that fails keeps its passage, so a total TypeSafe outage returns exactly the results we would have had anyway. Worse ranking, never a broken search. But it logs every degraded request, loudly, because a reranker that has quietly stopped reranking looks exactly like one that is working.
It also decides which messages reach the tools at all. Our assistant has 72 tools wired into it, and the agent that holds them is a 76,000-token proposition before anyone has typed anything. So Jev now takes one yes/no question about each message first — does this need our private data or our devices? — and answers it in about 650 milliseconds for roughly 340 tokens. Anything scoring 0.25 or below never wakes the agent at all. Which is exactly the use Ronacher pointed at the week the model shipped.