Which model drives a desktop best

If you are about to point a computer-use agent at a Linux desktop, the model matters more than anything else you will choose. On the one public dataset not produced by a driver's own authors, the Claude models pass about half again as often as the GPT models on the same tasks, with the same driver. These are the numbers, recomputed from the per-trial data on 2026-11-20, and the caveats that come with them.

Scale AI's dataset, recomputed 2026-11-205,370 scored trialsNot our benchmarkUpdated 2026-11-20

Where these numbers come from

The honest provenance, because benchmark tables are usually run by whoever wins them.

Cua Driver Bench v2 is ten desktop tasks, five runs each, driven in Claude Code for the Claude models and Codex CLI for the GPT ones, with 45 turns and 360 seconds allowed a trial. Scale AI published the per-trial dataset for it. We recomputed that dataset ourselves rather than quoting a chart: one row per operating system, condition, driver, harness and model, which is how the model table below exists at all.

  • Scale's run is the only published one not made by a driver's own team. Cua writes the tasks and the checkers and tunes its driver on them.
  • Every other leaderboard we found is run by the maker of one of the drivers in it, and each says so.
  • Nobody accepts outside submissions, so a new driver's number is always its own author's run.
  • This is not a Kubi benchmark. We have no arm in it, and when we run one we will say what it measured.

Linux, one model at a time

Out of 50 trials each, on the foreground condition: the desktop belongs to the agent and nobody else is typing.

Modelcua-driveragent-desktop
Claude Fable 5.24340
Claude Sonnet 5.54239
Claude Opus 5.53837
Claude Haiku 5.53437
GPT-6.1 Sol2630
GPT-6.1 Terra2630
GPT-6 Luna2629
GPT-6 Astra2530

Scale AI's per-trial data for Cua Driver Bench v2, recomputed 2026-11-20. Fifty trials a cell.

Two things stand out. The spread between the best and worst model is wider than the spread between the two drivers, so the model is the choice that matters. And the GPT models trail here partly for a plumbing reason Scale names: Codex CLI passes the driver's structured results on without their screenshots, so the model is working with less than the Claude side gets.

Price moves the other way from what you would guess. On Linux with Sonnet 5.5 the mean model cost a trial was $0.598 through cua-driver and $0.405 through agent-desktop, for a 42 and 39 out of 50. A cheaper model that fails more often is rarely cheaper once you count the retries, and the machine underneath is a flat $20 a month either way.

What happens when a person is using the machine

The background condition: the same tasks, with a simulated person typing in front of the agent.

Operating systemDriverPassRate
Linuxcua-driver40/32012.5%
Linuxagent-desktop37/32011.6%
Windowscua-driver61/24025.4%
Windowsagent-desktop0/2400.0%
macOSAside64/8080.0%
macOScua-driver45/8056.2%

The same dataset, background condition. A Linux result near 12% is not a driver that broke; it is what sharing one desktop does.

From about two thirds in the foreground to about one eighth in the background is the single largest effect in the data, and it is not really about models at all. One desktop has one mouse and one keyboard, and two parties using them at once is a race. On the Linux box used (Ubuntu 24.04 with XFCE) part of it is even more specific: xfwm4 hands the keyboard to every new modal dialog, so three of the ten tasks cannot pass there for any driver or model.

That is the argument for a lock rather than a shared mouse, which is how a Kubi screen works: when a person presses Take control, the agent's calls get a clear 409 until they release it, instead of both of them typing into the same window. It is also why the window manager on the image is worth caring about, and why we run GNOME rather than XFCE.

What to do with this

  • Start with a Claude model for a Linux desktop, and measure your own task before believing any table, including this one.
  • Hold the model fixed when you compare drivers. Most published comparisons do not, which is why they disagree.
  • Do not let a person work on the agent's desktop while it runs. Give each its own screen, or use a lock.
  • Count cost per pass, not cost per trial: $0.598 a trial at 84% beats $0.405 at 60%.
  • Read the interval, not the number. Fifty trials a cell means a few points of difference is noise.

Questions people ask

What is the best model for computer use on Linux?

On this dataset, Claude Fable 5.2 at 43 of 50 and Claude Sonnet 5.5 at 42, with Opus 5.5 at 38 and Haiku 5.5 at 34. The four GPT-6 models scored 25 to 30. Your own task is the only benchmark that decides your case, and a screen is the cheapest place to run it.

Does Kubi make me use a particular model?

No. Kubi rents the machine; the agent and the model are yours, and the API and MCP server are plain tool calls with an image, so anything that reads images can drive a screen.

Why do published computer-use leaderboards disagree so much?

Because almost all of them are run by the authors of one of the drivers being compared, on tasks they wrote, sometimes with the model varying between rows. Holding the model fixed and using somebody else's tasks changes the ordering, which is why we recomputed a dataset rather than quoting a blog.

Give your agent a desktop that stays on.

Private beta. Founding price $20 a screen a month, kept while you stay.

Get early access