A computer server circuit board from a data center

Paper explainer · preprint · 3 Nov 2026

A small AI learned to use a computer. The secret was a terminal.

Researchers report one 27 billion setting model jumps from 6.6 to 43.9 on a hard test once it may type commands. Here is what that costs, and what it does not prove yet.

vizmaya desk · 3 November 2026

The puzzle

Same brain. Two very different scores.

6.6
Mouse and screen only · WeaveBench score out of 100
43.9
Mouse, screen and a terminal · same model, same test

Both numbers are the authors' own report. Both come from the same 27 billion setting model, called CUA-Master. The only change is what it is allowed to do.

So what is a terminal doing that a mouse cannot?

From first principles

What is an AI that uses a computer?

It looks at a screenshot. It decides where to click or what to type. Then it looks again. It repeats until the job is done.

Researchers call this a computer-use agent. Think of a new office assistant who can only see the screen and has to do everything with a mouse.

A person working on a laptop away from a desk
Photo: "Mobile Worker" by mikecogh / Flickr, CC BY-SA 2.0, via Openverse. An AI agent does what this person does: look, click, type.

A mouse is slow for some jobs. Renaming 200 files by clicking is painful. A person who knows one typed command does it in a second. This paper asks what happens when the AI gets that same shortcut.

The claim

The headline: 80.2 out of 100.

80.2

That is the authors' score for CUA-Master-27B on OSWorld Verified. It is a test of 361 everyday computer jobs, such as office work, settings and file tasks (the paper's Table 4). Higher is better.

In plain terms: on the average job, it gets about four fifths of the points. It does this with a screen-only setup, no terminal.

PaperCUA-Master: Scaling Computer-use Agents through Open-World Recursive Improvements
AuthorsXinlei Yu, Gen Li, Ji Zhao and 15 others; corresponding author Xiangyu Yue
GroupsMMLab CUHK, THU, HKUST(GZ), Explore Academy JD
IDarXiv:2611.03126 · 2 Nov 2026 · preprint not peer reviewed
Linkarxiv.org/abs/2611.03126

Is 80.2 a big deal? Depends who is standing next to it.

In context

On the headline test it comes seventh of twenty.

Six models score higher, including a same-size rival from another team, Qwen-3.8-27B, at 81.2. The authors call this "comparable". That is a fair word for one point.

Seventh of 20 on OSWorld Verified, in the authors' own table

OSWorld Verified · average score out of 100 · authors re-ran every model under one setup · preprint
Source: arXiv:2611.03126, Table 1, v1, 2 Nov 2026. Closed models were called through their public APIs. Single run per score, no error bars given.

If it is only seventh here, why is anyone writing about it?

The twist

Use both tools and the score jumps.

Two tests mix screen work and command work. On these, the mouse alone is weak. The terminal alone is weak too. Together they are far better than the sum would suggest.

Mouse plus terminal beat either one alone

Same CUA-Master-27B · WeaveBench average score · Agents' Last Exam full-pass rate · preprint
Mouse onlyTerminal onlyBoth
Source: arXiv:2611.03126, Sec 3.3.1 and Fig 6. The terminal-only figure for Agents' Last Exam (18.4) is read from the figure's labels, not the text.
23.8%
fewer steps with both tools than mouse only, on WeaveBench
26.7%
fewer steps on Agents' Last Exam

Better and shorter? How does a terminal save steps?

Show, don't tell

The agent swaps tools mid-job, like a person would.

The paper shows three jobs where it switched back and forth. The terminal does the bulk work, such as reading files and checking answers in one go. The screen does what needs eyes, such as filling a web form or inspecting a brain scan.

In three showcased jobs, the share of terminal steps ran from a third to three fifths

Share of steps by tool · three examples picked by the authors, not an average · preprint
Screen stepsTerminal steps
Source: arXiv:2611.03126, Fig 5. Jobs: a web application from two Word files (40 steps), grading answer sheets into a spreadsheet (61 steps), checking brain scan masks (102 steps).

Three hand-picked examples show it can happen. They do not show how often it happens.

Fine, but where did it learn this? Nobody wrote 127,000 lessons by hand.

Where the lessons come from

It learned from the internet's how-to videos.

People post screen recordings that show how to use software. The team collected them for more than 650 apps, across 8 areas of work and life.

  1. 1. Find the videosSearch YouTube, TikTok and Bilibili for tutorials on each app. Throw out ads and talking heads.
  2. 2. Cut each video into jobsOne tutorial can hold several separate tasks. A larger AI model marks where each one starts and ends.
  3. 3. Turn the demo into clicksMatch what the speaker says with what changes on screen to write down the actions.
  4. 4. Remix the skillsBreak each job into small reusable skills, then combine them into new, longer jobs with a built-in answer checker.
  5. 5. Practice in real virtual computersThe agent tries each job in a virtual machine. If it slips, the system rewinds to the last good moment instead of starting over.

The result is 127,000 practice jobs: 115,000 to copy, and 12,000 for trial and error with a scorer. The team then repeated the loop, using what the model got wrong to choose the next batch (the paper's Sec 2.2 and App B).

Step 5 sounds small. Is it?

The quiet trick

Rewinding beats restarting.

Long jobs fail somewhere in the middle. If you restart every time, you almost never finish. If you rewind to the last good step, you keep most of the work.

On the longest jobs, restarting almost never works and rewinding sometimes does

Share of practice runs that finish · K = skills in a job, N = apps involved · authors' own baseline · preprint
Restart from scratchRewind (CUA-Master)
Source: arXiv:2611.03126, Table 3. Baseline: same retry budget, naive environments, restart from the beginning. It is the authors' choice of baseline.

More runs that finish means more good lessons. And more lessons means more of what?

Lens: what it costs

The chips are small. The bill is somewhere else.

128
H200 chips used for training: 16 servers of 8
89.6 kW
chip power at full load, my arithmetic (128 x 0.7 kW)
2,000+
virtual computers running practice jobs at once
1 server
to run a model of this size, against 6 for Kimi-K3
Rows of server equipment in a data center
Photo: "Data Center Servers" by Cory M. Grenier / Flickr, CC BY-SA 2.0, via Openverse. Illustration only, not the hardware in the paper.

Size does not decide the score: 27B models sit from 54.6 to 81.2

OSWorld Verified score vs total parameters (billions, log scale) · bigger dot is this paper · Qwen-3.8-27B sits just above it at 81.2 · preprint
Source: arXiv:2611.03126, Table 1 and Sec 3.1.1 (parameter counts from the model names listed there). Mixture-of-experts models use only part of their parameters per word. This is a size proxy, not training compute: the paper does not report FLOP, GPU-hours or dollars.

Lens read

Training took 128 H200 chips, which draw at most 0.7 kW each (NVIDIA). That is about 90 kW for the chips alone, with no cooling or network counted. The paper does not state FLOP, GPU-hours or dollars, so no total can be checked.

The authors name the real costs themselves: the virtual computer fleet, trajectory collection, and a large model to label videos. Stronger labelers gave bigger gains (Fig 9, App E). The saving shows up at use: a model this size runs on one 8-chip server.

If it is that cheap, what has to be true for it to matter?

Beside it

A cousin with a different job.

The closest paper in AI Daily's pool is Self-Credit (2610.28388, 20 October 2026). It teaches web-search agents which step in a long run deserved the reward. The authors report a 37.0 point gain on a test called BrowseComp-Plus.

The overlap is the last stage of training. Both use trial and error with a scorer. The difference is where the lessons come from. Self-Credit improves how credit is shared inside the model's own tries. CUA-Master brings in new lessons from outside, in the form of tutorial videos. CUA-Master does not cite Self-Credit.

Both are preprints. Neither has been repeated by anyone else.

What it would take

How we will know if this holds up.

It is two days old. Every score in it is the authors' own, from one run each. Here is what would settle it.

The thing to watch is not the biggest model. It is which tools you hand the model you already have.

Go further

Sources and claims log

Paper: arXiv:2611.03126, CUA-Master: Scaling Computer-use Agents through Open-World Recursive Improvements (Yu, Li et al.), arXiv preprint, 2026-11-02, https://arxiv.org/abs/2611.03126. Picked from AI Daily's papers (https://vizmaya.fyi/ai-daily/doom-v-boom).

Claims log summary

17 claims checked. 14 verified, 11 of those as claims from the one preprint. 1 contested (release status). 1 contradicted (affiliations in AI Daily's early reading were wrong; the paper lists CUHK, THU, HKUST(GZ) and JD). 1 unverified (a figure of about 10^24 FLOP, which the paper does not state and which does not fit a 128 chip run).

Not found: training FLOP, GPU-hours or dollars; a contamination check; error bars; a peer-reviewed version. OpenAlex has no record yet, so no citation count.

Photo credits

vizmaya desk · 3 November 2026 · Every score in this story is from a preprint and is reported as the authors' claim.