Routing three Claude Code loop decisions to Jev cuts monthly bill from $765 to $3.02
In the same nightly loop of 600 decisions, moving the questions of whether there is anything to fix, whether a command is safe and whether the goal is done to Jev's Noul and Choice cut cost from $765 to $3.02.
13.7 seconds. Let me pull frames.$765 a month turned into $3.02 after 1 change to a Claude Code loop. Same loop. Same 600 decisions a night. Same PRs waiting in the morning. The only thing that moved: who answers the small questions. An overnight agent asks 3 questions over and over: Is there anything to fix? Is this command safe? Is the goal done? Every one goes to Claude. Every one ships ~4K tokens of state. Every one costs ~$0.04. 600 a night. 30 nights. $765 for answers that are mostly "no." Part 1. The model that doesn't write On September 15, TypeSafe AI came out of 2 years in stealth with a $40M seed led by DCVC. The founder, Diogo Almeida, co-invented RLHF at OpenAI. His model, Jev, doesn't generate text. You send state and a typed question. It sends back a calibrated probability. 3 question types: Noul: yes or no, returns 0 to 1 Choice: picks 1 of up to 255 options Score: places the input on your scale $0.042 per 1M input tokens. Output is free. 70 to 500 ms per answer. Part 2. The split Triggers stay the same. A 5-minute schedule. A new commit. A new issue. A changed file. 3 gates move to Jev: Triage → Noul: anything to fix? Safety gate → Choice: allow / deny / ask Stop condition → Noul: goal complete? Claude only wakes up for real work. Fix the code. Run the tests. Open the PR with an explanation. Nothing to do? The loop sleeps. Claude never gets called. Part 3. The bill Claude decides: $0.04 per decision → ~$765 a month Jev decides: $0.00016 per decision → $3.02 Jev, batched questions: $0.00005 → $1.08 ~100 ms per decision. Before: Claude reads 4,000 tokens to say nothing changed. After: Jev returns 1 number, and Claude stays asleep. TypeSafe's own evals put Jev at ~445x cheaper than frontier models on decision tasks. Access is waitlist-only. Link in reply. The model is named after William Stanley Jevons. His paradox: make something cheaper, and people use far more of it. Your agent was paying $765 a month to say no. If this was useful - follow.
This post quotes another post: x.com/polydao/status/2103689373774483815