Summary
Creator Magic demonstrates Jev, a fast, ultra-cheap classification model from TypeSafe AI built by a former OpenAI researcher who co-authored the original ChatGPT paper. Unlike a chat-based frontier model, Jev makes structured decisions in about 200 milliseconds and was used to sort roughly 250 real inbox emails for just over one cent, compared to Claude Opus 5 taking nearly seven minutes and costing $1.39 to do the same task.
The video shows how Jev works inside the TypeSafe console, running categorization, true/false evaluations, confidence scoring, and multiple-choice classification on live email examples, including spam and even a simulated prompt-injection attack asking the model to run a destructive “rm -rf” command. Because Jev can’t generate free-form text or take direct action, it can’t be hijacked the way agentic models can, making it a lightweight safety filter that only escalates emails needing a real response to Claude.
The creator walks through the full automation stack: a Zapier zap watching Gmail for new messages, webhooks posting email data to a vibe-coded app, and Jev triaging everything into categories like sponsor leads, customers, and finance. The video closes with a look at Claude’s Small Business plugin and its Zapier connector, showing how cheap classification models and frontier models can work together in a practical inbox-management workflow.
📺 Source: Creator Magic · Published September 21, 2026
🏷️ Format: Hands On Build
