Jev, a cheap fast decision model
- YT :: https://www.youtube.com/watch?v=BHPDsGVciDk
- Original title :: This Might Be the Best AI Release of 2026
Prime is excited about Jev, a model that is not a chatbot but a decision engine: you hand it a state and a set of questions with constrained answers, and it returns classifications with confidence scores in about 100 ms for roughly $42 per billion input tokens. He frames it as "system 1 thinking" for software — the reflex that dodges a ball rather than the deliberation that writes an essay — and argues that being that cheap and fast unlocks a class of problems nobody bothers to solve with LLMs today.
What Jev actually is
Jev's own manifesto pitches "composable AI, build prod not god", with an analogy about early cars carrying whip holders because designers copied horses. The claim is that the industry is bolting chat interfaces onto problems that are really decision problems. Jev only does decisions: it takes a state description, then a batch of questions that run in parallel. Answer types include a choice among enumerated options, and a "new" type (apparently after the Bernoulli distribution) yielding a value between zero and one. Results come with confidence or relative scores. You cannot converse with it and you cannot ask it to write quicksort. The vendor claims roughly 190x faster and 444x cheaper than an unnamed baseline model.
The demos
Prime walks through toy prompts — a scatological "what should I do" choice, and a Super Smash Bros. moderation case asking both "is this message inappropriate" and "does this help Donkey Kong win". The second maps onto a live event he ran where 2,000 people drove four characters via chat: filtering messages through a conventional agent took about ninety seconds per round, forcing him to stall on stage. Per-message parallel classification would have collapsed that to nothing.
Beyond moderation, he shows Jev playing Balatro in real time — it commits to a strategy (pair scaling), picks jokers to match, and wins roughly 30% of runs, while Prime loudly disagrees with its play. Other demos: diagram generation in real time, Mario, and Doom, where the trick is answering several parallel binary questions per frame (dodge or not, fire or hold). Computer use is the one he finds most compelling for work: form filling and UI verification are classification problems, so this could replace brittle Playwright tests with "go do these activities and confirm they worked".
Why cheapness is the point
The name nods to Jevons paradox: the cheaper it gets, the more of it gets used. Prime burned 8 million tokens playing Balatro and barely moved the bill; the Smash Bros. question at 345 input tokens works out to roughly three million invocations for $44, and you are only charged for input tokens. That price point makes ideas viable that would be absurd at chatbot rates — categorizing email urgency inline as messages arrive, or pruning an agent's context by asking "is this chunk relevant to the job at hand" and discarding what is not.
Takeaway
His enthusiasm is explicitly contrasted with model-release fatigue: Sol 6 and Opus 5.5 both landed the same day and he had not touched either, because they are more of the same. Jev is interesting to him precisely because it is a different shape of tool — narrow, fast, cheap, and fun to build with, rather than another attempt at general intelligence. Notably, he does not think you need a new language for this; an ordinary conditional calling out to the model would do.