Computer use built on Jev

Watch out

This is a build log thread, not an evaluation. The cost and latency figures are per step as quoted in the post, with no step count, no task suite, no success rate and no token accounting; the repository the thread links carries its own tagline of about $0.0002 a step, a different number from the post, and neither figure states what a step includes. The headline multiples are a comparison against Opus 5 on this one loop, not general model benchmarks, and the author's own framing is Jev as an aid for LLMs rather than a replacement. OCR was chosen for operating-system coverage, and the post concedes that not all apps or operating systems expose accessibility trees while arguing those trees would be faster and more performant in most cases. Not every step is a Jev decision: when no element is clickable the loop hands control to small language models, so a cost per completed task would mix both. An open question in the thread, whether a browser DOM processor would work instead of OCR, is not answered in the captured text, and no prompt, question set or code appears there. Copy reviewed against the linked page on 2026-10-04.