A model that only decides
On Sept 15, TypeSafe launched Jev, the first of what they call System One models. Jev does not write text. You send it a document, a customer record, or any text. You ask it yes-no questions, multiple-choice questions, or rating questions. It returns a number between 0 and 1 for each, plus a confidence score. That is all it does.
Jev costs $0.042 per million input tokens. Output is free. TypeSafe says that is 100 to 400 times cheaper than a frontier language model for the same classification task. One independent test ran 21 questions across 37 documents in 0.7 seconds.
What you can hand to it this week
Spam detection. Customer support triage. Flagging urgent emails. Scoring product reviews by sentiment. Ranking job applications by fit to a rubric. Labeling invoices by type. Any task you can express as a decision, a rating, or a yes-no answer.
Jev evaluates all your questions in parallel, so sending 20 questions costs about the same time and money as sending one. If you currently send 100 items to ChatGPT one at a time for classification, you can send all 100 to Jev with 10 questions each in a single call.
What you must keep in your own hands
Jev returns only a floating-point number. It does not explain why. If you use it to rank job applicants and it scores one candidate low, you will not know which part of their application caused that score. You cannot ask it to justify the decision.
This makes Jev dangerous for any decision where you must explain or defend the outcome. Use it where a wrong score costs you a few minutes to fix, not where a wrong score costs someone a job or loan. Test it heavily on your own data before trusting it in production. And do not use it for hiring, lending, or any decision where hidden bias could cause harm you cannot see.