The old story of King Solomon begins with two women, one child, and two incompatible accounts of whose child it is. Solomon could ask for more testimony, examine motives, wait for further evidence. Instead he shrinks the problem: propose to cut the child in two, then watch which woman objects. The test isn't the verdict itself, just the smallest device that could force the truth into view.
Something close to that instinct just showed up in software. On September 15, TypeSafe AI released Jev, the first of what it calls "System One Models," a name borrowed from Daniel Kahneman's Thinking, Fast and Slow: System 1 is the fast, intuitive judgment that answers before you've reasoned it through; System 2 is the slow, deliberate kind that shows its work. Jev doesn't generate text. Feed it unstructured input and a schema you defined in advance, and it returns a typed answer with a calibrated confidence score, trained for exactly that through what TypeSafe calls Reinforcement Learning for Calibrated Decisions (RLCD)—aimed at the decisions software makes constantly and never talks about: classify this ticket, route this call, flag this transaction, extract this field. TypeSafe's own benchmarks—the only kind it says it will publish—claim it at nearly 200 times the speed and over 400 times cheaper than asking a frontier model to do the same job.
The market noticed immediately: Jev became, within a single day, the fastest-adopted model in Vercel's AI Gateway's history; within a week, an independent registry called JevMade had catalogued over 900 experiments built on it. But the stronger tell came a little later, when a small flotilla of variants and lookalikes started showing up on their own—a stronger signal than any adoption chart, suggesting TypeSafe had simply named a category people were already reaching for.
Two Speeds, Recovered
Each step since 2022 replayed the same move: not a new kind of intelligence, but a new interface onto the same one. Conversation made intelligence something you could simply ask. Tool calls let it reach outside its own text to touch a database or an API. Multimodality gave it sight, so the world no longer had to be typed out first. Reasoning models turned computation into something a system could spend more or less of, task by task. Then came agents that acted, watched, and tried again, and open protocols that let capability travel between systems instead of staying locked behind one API.
Before any of that, BERT and the classifiers that followed it already did this well and cheaply. An encoder trained to flag spam or route a support ticket ran fast and cost almost nothing—it just couldn't do anything else—with a fixed menu of possible answers.
A generative model doesn't have that ceiling. It can answer open-ended questions, because it was never trained on a fixed menu to begin with. That freedom is also the cost: every answer, including the trivial ones, comes from a system built to write essays and refactor code. Jev's trick is to keep the old menu but stop welding it to the weights—the schema arrives with the request, not with retraining. It recovers the encoder era's cheapness without giving back the generative era's programmability, and the result is closer to a new category than to either parent.
A Decision Model recovers the encoder era's cheapness without giving back the generative era's programmability.
That's not the same move as structured output or tools—introduced as early as 2024—which still ask a generative model to speak, only inside a shape someone drew in advance. A decision model doesn't dress an answer in a schema. Choosing among the options is the task—there's no sentence underneath waiting to be constrained.
A Pattern, Not a Product
Within days, a flurry of imitators had names. SemIf, formerly called OpenJev—despite no affiliation with TypeSafe—takes existing generative models like Qwen and prunes them down to the same trick. Laya goes the other way, built on ModernBERT-large, a classifier from the old encoder lineage rather than a repurposed generative one—open-weight and small enough to run locally, without depending on anyone's API. Google's own Gemma team showed that DiffusionGemma, a model built for entirely different purposes (like Laya), could be pointed at the same task by evaluating every option in parallel through diffusion instead of one token at a time.
None of these are Jev, but together—part of a wider ecosystem of forks and rewrites—they read less like competitors than like the paper trail of an idea whose time had already come: the decision model as a category of its own, not a truncated chatbot.
A Tax on Intelligence
A week later, Anthropic released Opus 5.5 at roughly 40% less per typical task once faster inference and lower token counts were counted in. The same day, OpenAI cut its own prices too, by up to half, with GPT-6 Sol and Luna!
Anthropic and OpenAI are cutting the cost of intelligence: same task, a cheaper frontier model, fewer tokens to reach the answer. Jev's entire category is asking a deeper question upstream of that one: not which model is smart enough for the job, but how much intelligence the decision actually deserves. A decision harness routing to Jev isn't just cheaper—it removes degrees of freedom the task never needed, exactly what Solomon did to the courtroom.
For a given task, every unit of intelligence invoked beyond what the task needs is a tax. Paid most visibly in tokens and latency, but also in variance, in the extra verification a fluent wrong answer demands, in the governance a bigger model quietly drags along for a decision that never needed one. TypeSafe itself named Jev after William Stanley Jevons, the economist whose paradox says efficiency rarely shrinks consumption—it multiplies it, cheap enough to use everywhere. Make deciding cheap enough, and there will be more deciding, not less. The next efficiency frontier may not be cheaper intelligence. It may be intelligence a system never had to invoke.
* * *
PS. There's an uncomfortable mirror here for how large organizations actually work. They rarely suffer from a shortage of intelligence—committees, expertise, analysis mobilized around a question are rarely in short supply. What's more often missing is a decision interface, not one more layer of thinking: a question bounded enough to answer, named alternatives, someone accountable for choosing, a moment where the thinking has to stop.
Deciding gets treated, too often, as a privilege of rank—the reward for climbing high enough to be the one who chooses. It works better read as a service: every day a decision doesn't land is a cost paid by people who aren't in the room to make it. Solomon wasn't short on advisers. His task was never to out-think the room. It was to make the problem small enough to decide—not for his own sake, but for two women who couldn't decide it themselves.
