Jev reads 37 documents and answers 21 questions in 0.7 seconds
In a test, Jev turned questions into yes/no, multiple-choice and scoring tasks, processed 37 documents in 0.7 seconds, and was 25x faster and 580x cheaper than Fable 5.1, though it missed one error.
AN AI READ 37 DOCUMENTS AND ANSWERED 21 QUESTIONS IN 0.7 SECONDS. THE BILL: A QUARTER OF A CENT. FABLE 5.1 COSTS 580× MORE. and it missed exactly one f*cking error. stay for that part, it's the whole point ok so here's what they actually did: model's called Jev. TypeSafe AI, ex-OpenAI founder, came out sep 15 Every gave it 37 docs and 21 questions pink boxes in the video are Jev > reads all 37 docs > turns each question into yes/no, pick one, or a score > doesn't write anything back. just the answer and how sure it is > done in 0.7s. i had to slow the video down like 45x so you can actually see it now watch the right side at 0:33. Jev hits DONE and Fable 5.1 is sitting at 4% that's the side by side Every ran on spotting errors Jev caught 6 of 7. Fable caught 7 of 7 Jev was 25x faster and 580x cheaper so yeah. it missed one i did the math. 1,000 runs a day = ~$2.50 on Jev. ~$1,450 on Fable. every day want to try it in 2 min: > it's on OpenRouter, your normal key works > ask it yes/no stuff you check by hand every day > look at the "how sure" number. anything under 0.7, send to a bigger model paying 580x more for 1 extra error isn't quality. it's a habit
This post quotes another post: x.com/hanakoxbt/status/2101709924828934222