Jev vs GPT-5.5 and Opus 4.8: 15-run content moderation consistency test
One borderline post is run through 8 questions 15 times to see whether each model flips queues, comparing per-call prices of $0.000046 and $0.041.
it's f*cking the best i found this weekend TypeSafe just made GPT-5.5, Claude Opus 4.8 and Jev moderate one post 15 times a borderline post can land in a different queue on the very next run. the test counts how often each answer holds: 1 post -> 8 questions -> 15 runs -> under 0.60 goes to a human all nine setups, drawn on two pages: > see which models flip the same post between queues > watch what a 0.60 floor does to every model's agreement > compare price and speed per call on one log chart $0.000046 vs $0.041 per 8-question call, Jev against GPT-5.5, at the cookbook's listed prices. the rare test that prints its own caveats on the chart
This post quotes another post: “TypeSafe's official skill teaches Claude Code to build on Jev”