Jev: The AI Model That Makes Decisions Instead of Writing Text

On September 15, 2026, a startup called TypeSafe AI shipped a model that works against the core of what language models are. Jev returns no prose at all. Given a piece of state and a set of declared questions, it answers with typed decisions: yes or no, one option picked from a fixed list, a value on a defined scale, each answer carrying a probability and a confidence score. The founder, Diogo Almeida, previously worked at OpenAI on the research behind RLHF and InstructGPT, the methods that later became ChatGPT. His diagnosis after four years: chat models write better than humans, yet the promised automation never followed, because software cannot reliably consume free-form text. TypeSafe calls its new model class System One Models, borrowing Daniel Kahneman's split between fast intuitive thinking and slow deliberate reasoning.

Under the hood sits a purpose-built stack: a new model architecture, a parallel sampler, and a training method named Reinforcement Learning for Calibrated Decisions, or RLCD. A language model generates token after token in sequence; Jev produces every answer for a request in a single parallel pass. The response space is fixed before the call: questions, options and scales are declared in advance, and according to the vendor the model therefore cannot commit type errors, because every possible answer is mathematically confined to the declared schema. Each answer ships with a calibrated probability, and calibrated here means measurable: higher confidence tracks higher accuracy. From the perspective of your code, Jev behaves like a function call with frontier intelligence attached: unstructured state in, typed probabilities out.

A hallucinated sentence in a chat is an annoyance. A hallucinated tool call inside an agent does real damage, above all in systems with latency guarantees or deep dependency chains. That is TypeSafe's argument for treating type safety as the entry ticket to automation: the answer space is fixed before the call, escaping the schema is claimed to be mathematically impossible, and a single counterexample would refute the claim. The subtler question is honesty about uncertainty. A model that solves a task 95 percent of the time can only automate it reliably if it knows when it belongs to the remaining five percent. The probabilities that ship with every answer exist for exactly that moment: below a confidence threshold the case goes to a human, above it the flow proceeds.

The latency figure is 70 to 500 milliseconds end to end. For contrast, the announcement cites benchmark roundups that put frontier language models at 3 to 329 seconds end to end, which is where the claimed factor of 40 to 200 for equal intelligence on System One tasks comes from. Input costs 0.042 dollars per million tokens, and output tokens are free. The homepage distills this into claims of 193.6x the speed and 444.6x the savings; the company itself frames both as the upper end of real-world gains. The measurement comes from TypeSafe's own workflow evaluations, in which every candidate model runs the same compute graph and the average of GPT-6 Astra and Fable 5.1 serves as the reference; there, Jev reportedly leads by close to two orders of magnitude. Two caveats sit in the same document: the measurements ran from laptops on the US west coast, and the pricing could be subsidized, with sustainability to be proven over time. Choice questions scale to 255 options, handled in two stages beyond that.

The comparison with chat models is deliberately apples to oranges: LLMs carry more context and generalize far beyond this task, while Jev gives up free-form generation entirely. The honest question is therefore about task shape, not raw intelligence. Anyone who runs the same classification, the same routing, or the same approval decision every week currently pays for a language model's ability to generate text, an ability this workload never uses. That is the gap TypeSafe targets with its claim of similar intelligence on System One tasks at two orders of magnitude less latency and cost. In practice, the declared questions, options and thresholds become part of your own code, which turns them into reviewable artifacts rather than prompt folklore.

The YouTube channel Sasch the Tasch spent three days with the model and showed what that feels like. He was still on the waitlist and ran everything through OpenRouter, where the context window caps at 32,000 tokens, so bigger batches need splitting. His setup runs an SEO agent and a CRO agent across several projects; both used to file every proposed change as a ticket waiting for a human yes or no, twelve proposals in the first measured week. With Jev in front as a pre-filter, the agents executed eight of the twelve directly while four went into his review queue as tickets, and he puts his own saved workload at roughly 70 percent. Case two was Google Search Console: instead of paying to analyze only his top 50 rows as before, he had the entire property evaluated, 19,943 decisions in two seconds for under one cent. Across three days he counted about 33,000 decisions in total, and building his first test cases with Claude Code from the OpenRouter example stayed under 15 cents.

The pattern predates the model: agents fail at autonomy not because they cannot write, but because nobody defined who decides. A calibrated yes-no verdict with a probability attached is precisely the part that goes between an agent and its next action when a gate has to stand there. TypeSafe positions Jev accordingly, as a fuzzy conditional for ordinary software: classify, route, score, extract, verify what a language model produced, flag jailbreaks, or compress petabytes into features map-reduce style. The limits are stated plainly: no text output, no explanations, a small context window, and pricing that is deliberately low while its long-term sustainability remains to be proven. For agent stacks this reshuffles the layers rather than replacing them: frontier models keep drafting and coding, cheap API models handle reactive chores, and decision models take over wherever the same question falls a thousand times a day. The name points at William Stanley Jevons: every order of magnitude by which intelligence becomes cheaper unlocks new work, the way coal demand grew once steam engines got efficient. The interface is small enough to wire up in an afternoon, which is the honest first step.

All links retrieved on September 23, 2026. Speed and pricing figures come from the TypeSafe announcement of September 15, 2026; practice numbers come from the video test cited above. The vendor's speed and cost figures are its own measurements and evaluations; an independent benchmark does not exist yet.

About the author: Marcel builds and runs HUNTER, an AI agent on a 45-euro Android cyberdeck, and writes about agents that run without a cloud engineering degree.