Testing Jev: a decision model that writes no text, accuracy and cost compared
The author reverse-engineers Jev and compares accuracy, per-case price and latency against GPT-5.6 Terra and Claude Opus 5 across four workflows, noting unverified points like self-run benchmarks and no rationale output.
I still can't believe a model this useful never writes text… i spent a day reverse-engineering Jev from the docs and hands-on tests what it does: > you send messy state + typed questions > it returns decisions with calibrated odds > your code routes on the confidence the numbers (TypeSafe's own eval, 4 workflows): → Jev: 67.8% accuracy, $0.0004 per case, 0.4 s → GPT-5.6 Terra: 67.9%, $0.0304, 10.1 s → Claude Opus 5: 73.1%, $0.1761, 37.8 s what's not proven yet: -> the benchmark was written by TypeSafe itself, no independent reproduction -> confidence is calibrated across many calls, not per single answer -> no rationale, so you can't audit why it decided so it's not a Claude replacement it's the layer in front of Claude that decides when Claude is even needed full breakdown on one page ↓
This post quotes another post: x.com/solst1ne/status/2101724066369654935