Computer Use, 1 month of testing. What works, what doesn't.

Anthropic pushed the latest iteration of Computer Use in early March. I used that as an excuse to seriously integrate it into my workflow for a month, on real tasks from my studio. Not Twitter demos. Actual work.

Short verdict: it’s useful on 3 categories, it’s smoke and mirrors on 4 others. Let’s break it down.

The setup

Claude Computer Use, March 2026 version, called through the API inside a homemade agent. Environment: dedicated Debian VM, 1920x1080 virtual screen, headless with controlled access to a subset of applications. I go through the API rather than Claude Desktop because I want to log everything that happens and replay sessions when something breaks.

The tasks I tested, real ones, pulled from my studio:

1. Extract data from a legacy analytics dashboard with no API
2. Manually test a payment flow on Stripe in test mode
3. Take annotated screenshots of an app for documentation
4. Fill out a tedious admin form on a client platform
5. Check the SEO of about ten pages (open, inspect, note)
6. Debug a frontend that works locally but not on staging
7. Extract info from scanned PDFs stored in a Drive

What worked

1. Extraction from an API-less UI (“light RPA” category)

The client’s analytics dashboard runs on PHP 5.6 that spits out raw HTML, no JSON endpoint, no CSV export, nothing. The only historical fix: a human clicking through the table, copying, pasting into Excel. 40 minutes, 3 times a week.

Computer Use does this in 4 minutes, 100% reliable across 15 consecutive tested sessions. The agent opens the dashboard, navigates, extracts, formats as CSV, commits the file to a backup repo. I set it up as a daily cron job.

Actual measured savings: 6 hours of human work eliminated per month. API cost: ~12 €/month.

2. Filling out tedious, repetitive forms

Typical case: an admin platform where you need to create user accounts with about twenty fields, step-by-step validation, a 4-page wizard. The agent does this reliably, better than a human even, because it doesn’t get tired by page 3 and doesn’t skip a field.

I use it to onboard new projects for a client who forces me through their portal. Gain: 20 minutes per new project that I no longer have to do.

3. Cross-device / cross-browser visual checks

The agent opens a URL, takes screenshots at different resolutions, compares against a baseline, flags discrepancies. It’s basically Playwright with a brain. For post-deploy visual checks on marketing pages, it gets the job done. Not as good as a real visual regression tool (Chromatic, Percy), but for a one-shot check it’s enough.

What didn’t work

1. Debugging a frontend — total failure

Task 6: a client’s Next.js site that works locally but not on staging. I asked the agent to open staging, identify the bug, suggest leads. Result:

• The agent didn’t open DevTools correctly. It spent 15 minutes trying to virtually click F12.
• Once it finally got there, it couldn’t read the console properly. It saw “a lot of red text” and reported that as “there are errors” without identifying them.
• It confused a React hydration warning with a fetch error.

Cost of the task: 2.30 € for an unusable diagnosis. I did it myself in 8 minutes.

Lesson: anything that requires finely reading technical output displayed graphically (DevTools, streamed logs, a complex IDE UI) is still far off. A CLI agent that reads files directly is a much better bet.

2. Extraction from scanned PDFs

I tried it on scanned PDFs (invoices, purchase orders). Failure: the agent can see the PDF, but the native OCR quality of its vision model on medium-quality scans (300dpi, tilted ink, stamps) remains worse than a good dedicated OCR tool (Tesseract 5 with preprocessing, or the Mistral OCR API). It hallucinates numbers, mixes up O’s and 0’s, misses fields.

For a human, no visible difference in the PDF. For the agent, 7% of fields were wrong on a batch of 50 documents. Not acceptable in production.

3. Manual testing of a payment flow

I wanted it to validate a Stripe test flow: create an account, add a card, place an order, check the confirmation email, cancel. The agent managed to get all the way to “place order” but:

• It filled fields too fast and triggered front-end validation that blocked it.
• It couldn’t figure out how to open Gmail to check the email (Google auth threw it into a panic).
• When the flow had an unexpected step (a cookie popup), it clicked on it at random.

For end-to-end testing, a well-written Playwright suite remains infinitely more reliable and cheaper.

4. Annotated screenshots for documentation

Worked on paper, but the annotations were… bad. Arrows in the wrong place, text overflowing, highlight colors that didn’t stand out. The result wasn’t publishable. I ended up taking the screenshots myself and annotating them in CleanShot in 2 minutes.

My verdict

Computer Use in 2026 is a tool that’s very good at light RPA (predictable navigation, data extraction, form filling), and frustrating at anything that requires fine interpretation of rich interfaces (DevTools, IDEs, complex dashboards).

Concretely, out of my 7 test tasks:

• 3 successes → integrated into my production workflow
• 4 failures → abandoned, I keep doing them the old way

The ROI on those 3 successes is real: roughly 8h/month saved on tedious tasks nobody wanted to do. At 20 €/month in API cost, it’s trivially profitable.

What I learned from this

Computer Use is not “an AI that replaces a human in front of a computer.” It’s a smarter RPA tool than the historical ones (UiPath, Automation Anywhere), but still far from a human generalist.

The gap between demos and practice is still huge. Anthropic’s videos show Computer Use buying plane tickets. In practice, the same task gets blocked by a Cloudflare captcha or a cookie popup. Demos are staged on friendly, predictable sites.

For pure dev work, the Claude Code CLI is still much more effective. Computer Use is only relevant when you have no CLI/API alternative. If the target system has an API, use it. If you can script it headless, do that. Computer Use is the last resort.

I keep using it, but within the narrow scope I identified above. The other cases got crossed off my todo list.