Skip to content
OpenDesktop
How it works
Privacy
Pricing
Enterprise
Blog
Download
Status:
on your machine
9:41 AM
blog
Notes from the workbench.
Benchmarks, engineering write-ups, and honest numbers.
2026-07-08
Four budget models. One harness. The whole benchmark.
OpenDesktop on Harness-Bench with four of the paper’s exact backends, all the way through the 106-task suite: GLM 5.1 beats the best published harness average, and three budget models land within half a point of each other.
read it