TypeSafe AI's Jev: decision model at $0.042 per million tokens

TypeSafe AI's Jev arrives as a deliberate break from the text-in, text-out LLM pattern. Instead of generating prose, it accepts a "state" object—a string, array, or name-value pairs describing a record—and returns typed probabilistic outputs: a confidence score for a statement, a probability distribution across choices, or a number along a provided scale. The company frames it as a "frontier-intelligence function call," and prices it unconventionally: input only, at $0.042 per million tokens, with output free.

The appeal is speed and cost for classification-shaped work. The article points to spam detection, labeling, prioritization, ranking, and search reranking as natural fits, where many questions can be evaluated in parallel against a single document. But the design also strips away even the illusion of explanation that text-generating LLMs offer. As the article puts it, if Jev marks something as spam, the only thing you get back is a floating point number, leaving which content signals drove the decision unclear.

That opacity is why the article singles out bias as a first-order concern, floating the example of ranking job applicants, where an unexplained score could conceal unseen bias that is experimentally tricky to pick apart. One informal experiment scoring Bay Area cities on "Good city?" ranked Cupertino top and East Palo Alto bottom—a result the author flags with a "Huh." The stakes, then, may hinge less on Jev's speed than on whether its cheapness (hundreds or thousands of experimental prompts for a few cents) makes rigorous evals practical enough to expose what its numbers are actually measuring.