Theo's computer use bench puts macOS last at 27%
An early, unpublished benchmark measures the desktop the agent drives rather than the model driving it.
Theo says he is building a benchmark that measures agent task success and failure rates across operating systems, and that in its current state macOS sits at the bottom with a 27% pass rate. Omarchy and t3os are at the top, both at 100%. He also teased that "The Nix fans are NOT gonna be happy about this one", without giving a number for it.
That is the whole disclosure. He has not published the task set, the agent or model used, the number of runs, or how a pass is scored. He framed the numbers as a current state of work in progress, not a result, which is the right amount of hedging for a one-line post and also the reason nobody should quote the 27% as a finding yet.
Why the number is interesting anyway
Computer use agents, meaning agents that drive a desktop through screenshots and synthetic mouse and keyboard input rather than through an API, are usually evaluated on the model. Swap the model, rerun the tasks, compare. The implicit assumption is that the operating system underneath is a constant.
Theo's framing inverts that. If the same agent passes every task on one desktop and fails roughly three quarters of them on another, the dominant variable is the environment, not the reasoning. Screenshot legibility, window management, permission prompts and accessibility APIs differ enormously between desktops, and any of those could plausibly be the thing a benchmark like this actually measures. Which of them it is, nobody knows yet, because Theo has not said.
What to do with it
Treat it as a hypothesis with a provocative number attached, not as a ranking. Two of the three named systems in the lead are tied to the person running the benchmark, t3os being his own, so a published task list matters more here than usual. A 100% score on anything nontrivial is also a signal that the tasks may be too easy rather than that the platform is perfect, and the same tasks producing 27% elsewhere makes the spread the interesting part rather than either endpoint.
If you are shipping anything that drives a real desktop, the usable takeaway is cheap to test yourself. Run your existing task suite on a second operating system before you conclude your agent got worse, because the failures may be coming from the window manager and not from the model. Theo's post is the only public detail so far.
Current state: MacOS at the bottom (27% pass rate). Omarchy and t3os at the top (100%).
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.