Decisions were never conversations

For three years we have been doing something that, spelled out slowly, is absurd. We had a system that needed to know whether an email was an invoice or a complaint. To find out, we summoned a machine trained to write essays, politely asked it to reply in JSON and nothing else, begged it not to add any commentary, then parsed whatever came back and hoped it had not opened with “Certainly! Here’s the JSON you requested.”

We called that putting AI into production.

On September 15th, a company that had spent two years in silence shipped a model that does not write. TypeSafe AI came out of stealth with a $40 million seed round led by DCVC, a$200 million valuation, and a product called Jev that does exactly one thing: it takes a state and returns a typed decision with its probability. No prose. No tokens emitted one after another. No friendly preamble.

Within 24 hours, 13% of Vercel’s paid teams had tried it. No previous launch on their AI Gateway had cleared 7% in a full day. The announcement drew 25 million views. The API buckled under demand.

The industry’s instant read was that we were looking at a new paradigm. A widely read newsletter called it “the ChatGPT moment for decisions” after a few hours of testing. TechCrunch ran with a headline about “a ChatGPT inventor” and a new kind of model.

It isn’t. And that confusion is more interesting than the model.

What Jev exposes is not an unprecedented architecture. It is an invoice. For three years we have been paying frontier prices for operations that were never conversations, and calling intelligence what was in fact eloquence bought by the pound to obtain a boolean.

We commissioned an essay to get a yes

The mechanism is worth spelling out, because that is where the waste lives.

An autoregressive language model generates left to right, one token at a time, each one conditioned on everything before it. That sequential dependency is what lets it write a coherent paragraph. It is also what makes it slow and expensive: you cannot parallelise a chain where every link waits on the previous one.

Now point that at classifying a support ticket. The useful answer is one word: technical. But the model does not know how to return a word. It knows how to return a sequence. So it produces the braces, the quotes, the field name, the colon, the value, the closing bracket. Ten, fifteen, thirty tokens of scaffolding to carry three bits of information. And every one of those output tokens bills, on frontier models, at roughly five times the price of input.

Worse, the scaffolding fails. Not because the model is stupid, but because emitting valid JSON is a different task from deciding well, and we were asking for both at once from the same tool. Hence the entire parallel industry we built around it: structured output modes, constrained grammars, validators, retry loops, whole libraries whose only job is to persuade a text generator that this time, please, it should not be creative.

Jev gives up string generation. That is the move. With no sequence to produce, it samples every output in parallel in a single query and answers in 70 to 500 milliseconds. It exposes three primitives — Choice to pick among defined options, Score to rate on an ordered scale, Noul for yes/no — and hands back something like this:

{"billing": 0.08, "technical": 0.85, "sales": 0.07}

with an attached confidence of 0.82.

It was trained purely on synthetic data using a method the company calls Reinforcement Learning for Calibrated Decisions, RLCD, in deliberate contrast to the RLHF its own founder helped invent. Diogo Almeida, the CEO, spent four years at OpenAI and is a co-author on InstructGPT, the work ChatGPT came out of. Alongside him are Sasha Sheng, a research engineer from Meta/FAIR, as COO, and Erik Gafni, founder of Ravel and an early hire at Invitae and Freenome, as CTO. The team comes from OpenAI, Google Brain, Meta, Stripe, Airbnb, Plaid and Docker, and works five days a week in an office by Embarcadero.

It is a roster that opens doors and headlines. It is also why the industry took zero seconds to mistake a well-designed product for an epistemological break.

The news is not the model, it’s the price

$0.042 per million input tokens. Output tokens free. “Too cheap to meter,” says the website.

That is the story.

That price does not describe a technical feat. It describes what the operation actually costs — the one we spent three years buying dressed up as something else. When a Vercel engineer measures Jev classifying safety 5 to 18 times faster than OpenAI’s small model while being more accurate, or an email company finds classification running 10 to 20 times cheaper than with Gemini, they are not discovering a new kind of intelligence. They are discovering how badly they were overpaying.

13% of Vercel’s teams in a day is not enthusiasm for an architecture. Nobody adopts an architecture in 24 hours. It is the speed at which an industry moves once somebody shows it that a substantial share of its AI bill was a formatting toll.

And here a piece of honesty the hype skipped: the savings do not come from the architecture alone. They come from having drastically shrunk the problem. A model that can only choose among options you handed it is a model that needs to know almost nothing a frontier model knows. It is cheaper because it does less. That takes nothing away from its merit. It takes away the epic.

”Zero hallucinations” is a category error

The line repeated most this week is that Jev cannot hallucinate. TechCrunch printed it flat. It is false, and the interesting part is that TypeSafe admits as much on its own blog: the 0% hallucination figure “is not empirical.”

What is guaranteed is that the output conforms to the schema. Give it billing, technical and sales, and Jev cannot invent legal. That is impossible by construction, in the same way a dropdown cannot return a value that is not in the list.

What is not guaranteed is that it is right. Jev can return billing at 0.91 when the correct answer was technical, and in any production system that is precisely what we call a bug. We have swapped the risk of a model inventing a field for the far older, far less documented risk of it being confidently wrong inside the right one.

Guaranteeing shape and guaranteeing truth are different things. Schema validators have guaranteed the first for decades and nobody thought to sell them as the end of falsehood.

The confusion spreading was not a journalistic accident. It is how the machine works: a precise technical claim — zero out-of-schema outputs — travels badly, and a round one — zero hallucinations — travels beautifully. In 48 hours, a specialised tool became a paradigm. Nobody quite lied. Each link in the chain simply rounded up a little.

The confidence score is where responsibility now lives

The genuinely new part of Jev is the part that generated the fewest headlines, and the only one that could change how systems get built: it returns its own uncertainty, and returns it calibrated.

A calibrated model is one whose 0.7 means something. Say 0.7 across a thousand cases and you should be right in roughly seven hundred. Language models, tuned with human feedback to sound convincing, are notoriously bad at this: they skew overconfident because overconfidence is what human raters rewarded. RLCD optimises for the opposite.

If it works, the practical consequence is enormous and deeply unglamorous: it lets you write the rule you could never write before. Above 0.9, automate. Between 0.6 and 0.9, escalate to a large model. Below that, to a person. Armin Ronacher, CTO of Earendil, summed it up by saying it is the only one that hands back a real probability, and that if something comes back at 50%, maybe that was a coin toss all along.

But look at where the important decision just landed.

Not in the model. In the threshold. Somebody — an engineer, in a hurry, on a Tuesday — types 0.9 into a config line, and that number decides what runs unsupervised and what gets looked at. The model no longer just gets things wrong: it now tells you how wrong it might be and hands you the responsibility for having ignored it. That is a more honest split. It is also a split that turns a magic constant into the place where the whole system’s liability lives.

And here is the problem: calibration is exactly the claim the vendor has not documented. There is no paper. No published reliability curves, no expected calibration error, no architecture details. The founder himself has acknowledged that calibration is measured at group level, not per answer. Some observers suspect a fine-tuned open-weights model underneath.

So other people measured it, and a week in, a pattern is showing. An independent test published on the 19th found Jev well calibrated on public benchmarks, with a calibration error between 0.024 and 0.032, and considerably worse on support tickets that looked like nothing it had seen: 0.107, more than four times the noise floor. On questions that required knowledge the model did not have, it was right 44.7% of the time while assigning its answers an average probability of 0.74.

That 0.74 is the number somebody was about to write a threshold on.

Another test, on two thousand phishing emails, put it at 62.6% accuracy against Claude Haiku’s 81.3% when asked a single question. Split into five atomic questions and combined with a regression fitted on a thousand labelled examples, it climbed to 95%. The lesson is not that Jev is good or bad. It is that its number only means something once you have checked it against your own data, which is exactly the work the price seemed to save you.

The vendor’s own figures do not help. The launch numbers do not agree with each other: the tweet says 20-200x faster and 40-400x cheaper, the homepage says 193.6x and 444.6x, the press release says “up to 100x.” Independent tests land between 5x and 25x depending on the comparison. TypeSafe itself concedes — in a transparency exercise that deserves credit — that its test workflows were built by its own capabilities team, that the demos use short, dense inputs favouring its approach, and that accuracy is measured against the average of other models rather than against ground truth. There are no named customers. No revenue. Access opened to everyone on the 20th, and two days later new signups had to be shut off because demand outran what they could serve.

So what we have is a probably excellent product whose central argument — calibration — is currently a promise that holds where the vendor measured and loosens where you are going to use it.

They named it Jev for a reason nobody wanted to read

The model is named after William Stanley Jevons, the economist who observed in 1865 that making coal more efficient did not reduce consumption. It multiplied it. Cheapen each unit and uses appear that were never worth it before.

TypeSafe says so without hedging: they expect cheap AI to be deployed far more, not less. It is an honest statement of intent, and it is also the part the enthusiasm skipped.

Because a decision at $0.042 per million, in 70 milliseconds, with no output cost, is not the same decision that used to cost a hundred times more. It is a decision made in places where no decision was made at all. When deciding is effectively free, deciding becomes the default: in loops over millions of records, on every log event, in every interaction of every user, in intermediate layers nobody reviews because nobody designed them to be reviewed. The official demo was Jev playing Doom from the game state. The Register, which covered the launch with healthy scepticism, closed its piece by mentioning drone targeting decisions.

You don’t have to go that far. The boring version is enough. A model whose output is not text is a model whose output does not get read. It leaves no transcript a human might skim out of curiosity, produces no explanation somebody might find odd. It leaves a number. And numbers, unlike paragraphs, get aggregated into a dashboard and taken as given.

We have spent three years worrying about AI writing false things, because writing is visible. What comes next decides quietly, very fast, very cheaply, and everywhere.

A layer, not a paradigm

Fairness is in order: Jev looks like an excellent product. It solves a real problem anyone who has shipped an LLM has suffered, it solves it with an elegant design decision — give up generating text — and it prices that solution in a way that reshuffles budgets. Its founders know what they are doing. The Vercel adoption is genuine, not bought.

But classification, scoring, routing and intent detection are old machine learning problems. Zero-shot classifiers have been doing something adjacent since 2019. That a specialised model beats a generalist on a narrow task is not a discovery: it has been the rule for decades. What is new here is a specific combination — non-autoregressive architecture, calibration-oriented training, careful developer experience, aggressive pricing — and specific combinations build excellent companies. They do not build paradigms.

What actually happened in September 2026 is not that a new form of artificial intelligence was born. It is that the industry admitted, by way of a price, that most of what it called conversation with a model was never a conversation. It was a query dressed as a chat, billed at chat rates, carrying a formatting risk that kept us busy enough not to look at the real one.

The costume is coming off now. What is underneath is a decision, a number between zero and one, and a threshold somebody will have to justify when something goes wrong.

It is not a new model. It is the end of a comfortable confusion.