Jev decision model classifies 500 emails for 3.5 cents
Jev only makes yes/no, option and urgency decisions in 70-500ms per call, and a test found it about 76x cheaper than frontier models at similar accuracy.
the guy who helped build chatgpt just raised $40M for a model that can't write a single sentence diogo almeida was at openai for chatgpt and instructgpt. two years in stealth. on sept 15 he came out with jev chatgpt taught models to talk. jev is physically incapable of prose. it does one thing: decides. yes or no. which option. how urgent 70-500ms per decision. $0.042 per million tokens now open your agent. every time it asks "is this source relevant?" or "which worker next?" it pays a frontier model to think out loud about a yes/no that's the leak one dev classified 500 emails for 3.5 cents. total. not per email another plugged it into a live monad order book. a real order every 300 milliseconds in their own test jev matched frontier-model consensus at 67.8% for $0.0004 a case. gpt-5.6 terra got 67.9% for $0.0304 same accuracy. about 76x cheaper the model that writes and the model that decides don't have to be the same model
This post quotes another post: x.com/danikkk_wqs/status/2101706326883655710