Ask a language model, the kind of AI that writes text, whether an invoice looks fraudulent. The answer comes back as a paragraph. The model weighs the line items, notes the vendor history and lands on a conclusion after roughly eight seconds of writing. A model from a startup called TypeSafe AI skips the paragraph. Fed the same question, it answers “clean” in about a tenth of a second, paired with a confidence score.
The startup calls it a System One model, after the fast, automatic kind of thinking psychologists describe. Its first version, called Jev, does not generate text at all. A language model builds an answer one word at a time. This one produces every part of the decision at once, then stops. The output is always a label, a score or a yes-or-no answer with a probability attached.
That limitation is the design. Jev only picks from a fixed set of pre-approved answers, so a wrong guess dressed up with false confidence has no way in. A language model reasoning freely sometimes wanders into a convincing but wrong explanation. A system that picks from a short list has nowhere to wander.
Classification systems in business software already pick from short lists to triage support tickets, sort incoming email and flag records for privacy review. Each one is built and trained for a single job. Jev takes plain text, a data record or a list of options, with no retraining for each new task. It returns a decision for whatever categories the question defines.
The pricing follows from the output. AI models read and write text in small chunks called tokens. TypeSafe charges about four cents for every million tokens the model reads, and nothing for what it produces. The most advanced language models charge twenty cents to ten dollars for the same input, plus more for their longer output. In the company’s own tests, the model ran 40 to 200 times faster than comparable top models, and one workflow test ran nearly 445 times cheaper.
TypeSafe released Jev as version 0.01, letting developers in from a waitlist. The model cannot explain its reasoning, write code or draft a memo, so a language model keeps that work. Jev takes the small, repeated choices, such as routing a ticket, flagging a transaction for review or ranking five options, the kind of choice most software runs on.
