Summary
Jev, a new AI model from the company Typesafe, never writes text. Instead it returns probabilities for answers chosen from a predetermined list, which makes it suited to classification, scoring, yes/no decisions, and routing. This video explains how it works using a support-ticket example in which Jev decides which team should handle a message, how urgent it is, and whether a person should review it.
The explainer compares Jev with chatbots like ChatGPT and Claude, which generate open-ended text step by step. Because Jev skips that stage and answers several questions at once, it runs much faster. Typesafe lists pricing at 4.2 cents per million input tokens, with answers free. The video walks through reported results: accuracy of about 67.8% for Jev against 67.9% for GPT 5.6 and 67.8% for Claude Sonnet 5 on Typesafe’s four-task test, and a 1,500-email classification test where Jev scored 96.4% against 97.5% and 98.5% for two Gemini models at a fraction of the cost. A Vercel engineer also reported large speed gains on unsafe-command detection, though without details of the benchmark.
The presenter also tests Jev on the number of Rs in strawberry, where it answered true for both two and three across 50 runs each. That shows constrained answers can still be wrong, so ambiguous cases should be escalated to a person or a stronger model.
📺 Source: Sabrina Ramonov 🍄 · Published November 17, 2026
🏷️ Format: Tutorial Demo







