· Solveion · Perspectives · 6 min read
Deciding is not writing
A new model from a ChatGPT co-inventor refuses to produce text at all. It only picks from options you defined in advance. The claim that it cannot hallucinate is true in a narrow sense, and the more useful point sits underneath it.
Something called Jev has taken over a lot of technical conversation in the past week, and the summary going around is that it is an AI that cannot hallucinate and runs a hundred times faster than ChatGPT. Both halves of that are defensible and both are easy to misread, so it is worth being careful.
Jev is the first release from TypeSafe AI, a company started roughly two years ago by Diogo Almeida, who worked on ChatGPT at OpenAI and on the reinforcement learning from human feedback that shaped it. It went to early access in September. The name comes from William Stanley Jevons, the Victorian economist, and that choice turns out to be the clearest statement of what the company thinks it is doing.
What it actually does
Every model most people have used produces text, one token after the next, and if you want a structured answer out of it you ask nicely and then parse whatever comes back.
Jev does not produce text. You give it some unstructured input, along with a list of the answers you are willing to accept, and it hands back one of those answers with a confidence number attached. The API offers three shapes: pick a category, give a score, or answer a yes-or-no with a probability. That is the whole surface.
It is also not a transformer. Rather than generating sequentially, it computes all its outputs in one parallel pass, which is where the speed comes from. TypeSafe trained it with something they call reinforcement learning for calibrated decisions, aimed at honest probability estimates rather than pleasant-sounding prose.
The framing they use is Kahneman’s. Language models are System 2, deliberate and slow and good at reasoning out loud. Jev is meant to be System 1, the fast judgement you make without narrating it.
The numbers, which are the reason for the noise
Latency lands between 70 and 500 milliseconds, against 3 to 329 seconds for frontier models on the same work. Input costs $0.042 per million tokens, around a forty-eighth of what a frontier model charges, and output is free because there is barely any. Per decision, TypeSafe puts it at roughly four hundredths of a cent, against three to eighteen cents elsewhere.
That is somewhere between seventy-five and four hundred times cheaper per decision. The demonstration doing the rounds is Jev playing Doom at about ten decisions a second, which is a strange thing to build and an effective way to make the latency point.
Why this matters more than it first appears
Here is the part we think businesses should sit with, and it has little to do with the benchmark table.
Look at what your AI systems actually do in production. Very little of it is writing. It is routing a support ticket to the right queue, flagging a transaction for review, scoring a lead, deciding whether a document needs a human, working out whether this customer message is angry enough to escalate. All of that is classification wearing a chat interface.
For three years the only tool available for that work has been a model built to write essays, so companies have been paying essay prices and essay latency for what is functionally a sorting decision. If the decision layer becomes its own product category, a great deal of current architecture is more expensive and slower than it needs to be.
The pattern TypeSafe is pointing at is a small fast model handling the routing, ranking and escalation, and a language model called only when something actually needs to be written or explained.
The hallucination claim, examined honestly
The marketing says Jev cannot hallucinate, and it reports a zero per cent structured output error rate. This is true and it is narrower than it sounds.
Jev cannot return an invalid answer because the set of valid answers is fixed before it runs. It cannot invent a citation or a tool name because it cannot write a citation or a tool name. The guarantee is about the shape of the output, not about whether the output is right.
A correctly typed wrong answer is still a wrong answer. Jev can confidently route a technical problem to billing, with a well-formed probability attached, and nothing about the architecture prevents that. One reviewer put it well: a confident wrong decision is not a hallucinated paragraph, but it is still a model error with operational consequences.
Worth knowing too that TypeSafe’s own chart showing zero hallucination is labelled as not empirical. It follows from the design rather than from testing, which is a reasonable thing to claim and a different thing from a measured result.
The accuracy picture
On TypeSafe’s own four-workflow benchmark, Jev scores 67.8 per cent. GPT-5.6 Terra scores 67.9. Claude Opus 5 scores 73.1 and GPT-5.6 Sol scores 74.1.
So it matches a mid-tier frontier model and sits five or six points behind the strongest ones, on a benchmark chosen by the vendor. There are no public benchmarks yet, no third-party accuracy numbers on anyone else’s domain work, and no production reliability data at all. For a model in early access from a two-year-old company, none of that is damning, but it does mean the honest summary is “promising and unproven” rather than “solved”.
What the name is telling you
Jevons is remembered for noticing that when coal-fired engines became more efficient, Britain burned more coal rather than less. Cheaper consumption produced more consumption.
Naming a cheap-decisions model after him is a thesis statement. The interesting consequence is not that your existing AI bill falls. It is that decisions you do not currently make become affordable. Checking every transaction rather than a sample. Scoring every document in the archive instead of the recent ones. Putting a judgement call inside a loop that runs ten times a second, which is what the Doom demo is really advertising.
That is the version of this worth planning around, and it is the same shape as every other cost collapse in computing.
What we would do about it now
Not migrate anything. It is early access, the accuracy is middling on the vendor’s own numbers, and the architecture is new enough that nobody outside the company has run it in anger.
What we would do is take an inventory. Go through your production AI calls and mark which ones end in a decision from a known set of options rather than in text a person reads. Our expectation is that the first number is much larger than most teams assume, and that it is the first thing anyone has quantified about their own systems in a while.
If that inventory is large, you have a cost and latency problem that exists today regardless of whether Jev turns out to be the answer. If it is small, this whole category is not for you, and you have learned that cheaply. Either result is more useful than an opinion about a model that has been available for a fortnight.