Ten tasks. That’s roughly how far ahead of the human baseline GPT-5.4 computer use landed, by our arithmetic, and it’s the number to hold onto while the launch coverage celebrates. OpenAI reports 75.0% on a desktop benchmark against 72.4% for people. Fine. Before you let it near your accounting package, run a small test of your own.
Key points
- What does the GPT-5.4 computer use score actually mean? It means GPT-5.4 completed a few percentage points more of a fixed set of desktop tasks than the human figure OpenAI quotes.
- Who were the humans it beat? It is not your staff. The same test says humans completed over 72.36% of its tasks, so people failed more than a quarter of them.
- Is the 83% work score any better? Slightly, and with its own catch. OpenAI says GPT-5.4 matched or beat professionals in 83.0% of comparisons on GDPval, up from 70.9% for GPT-5.2.
- What does a 25% failure rate cost you? More than the model bill. API pricing is $2.50 input and $15 output per million tokens, against $1.75 and $14 for GPT-5.2, and requests above the 272K window count double against usage limits.
What does the GPT-5.4 computer use score actually mean?
It means GPT-5.4 completed a few percentage points more of a fixed set of desktop tasks than the human figure OpenAI quotes. OpenAI reports 75.0% on OSWorld-Verified, up from 47.3% for GPT-5.2, against a human baseline of 72.4%. The gap is 2.6 points.
The OSWorld site describes 369 tasks in the original set, so the gap is roughly ten tasks, assuming the verified set is similar in size. That is our arithmetic, and an estimate. The jump from 47.3% is the real news. The lead over people is a rounding error with a press release attached.
Who were the humans it beat?
It is not your staff. The same test says humans completed over 72.36% of its tasks, so people failed more than a quarter of them. We don’t know who those participants were, but your bookkeeper with five years on your accounting package is unlikely to fail at that rate.
“Beats humans” is accurate for the test and misleading for your payroll.
OSWorld tasks span real web and desktop apps, file handling and workflows that cross several programs. A person new to those exact apps stumbles in ways your experienced team doesn’t. A model that matches the stumbling newcomer hasn’t matched the veteran.
Think of a driving test taken by people who’ve never seen the car. Passing it is worth knowing about. It doesn’t make anyone your delivery driver. The honest reading is that the model finishes three of four OSWorld tasks and misses the fourth, and a quarter is a lot when each miss might be a wrong invoice number.
Is the 83% work score any better?
Slightly, and with its own catch. OpenAI says GPT-5.4 matched or beat professionals in 83.0% of comparisons on GDPval, up from 70.9% for GPT-5.2. OpenAI describes GDPval as 44 occupations and 1,320 tasks, graded blind by experienced professionals who rank the AI and human deliverables.
Matching a professional on a deliverable once isn’t the same as doing it reliably every Tuesday. And “matched or beat” lumps outputs experts judged equal with outputs they judged better, so the figure is friendlier than “better 83% of the time.”
Every number here comes from OpenAI. It ran its evaluations at its highest reasoning setting in a research environment that, by its own note, may give slightly different output from production ChatGPT. Read 75.0% as a ceiling, not a forecast. Your office has odd software, old spreadsheets and logins that time out, and a test environment has none of that on the day.
What does a 25% failure rate cost you?
More than the model bill. API pricing is $2.50 input and $15 output per million tokens, against $1.75 and $14 for GPT-5.2, and requests above the 272K window count double against usage limits. On list prices that’s 43% more for input and 7% more for output, which is modest next to a 27.7-point jump in the score.
So the sticker isn’t the problem. The cost that surprises owners is repair time.
Here’s the shape, and it’s our illustration, not a measurement. A model does 100 desktop jobs a week at 75% success, so 25 need redoing. At four minutes a repair, that’s about 100 minutes weekly. Then add the harder cost, which is finding the errors that came back looking fine.
The same gap between a launch headline and a real saving turned up in our look at Sonnet 4.6 pricing. A model that can click through your desktop also needs permissions, which our piece on OpenAI’s consulting deals for AI agents treats as the hard part. If staff run agents on their own, our Moltbot security checklist is the practical starting point, and our note on the workplace AI adoption gap explains why leaders often underestimate how much staff already use.
How do you test it before you pay for it?
Test the failure, not the average. Pull 20 desktop tasks from your own last two weeks, such as moving figures between a spreadsheet and an accounting package, filing documents or updating a customer record. Run them in a throwaway account with copies of the data, and keep three counts.
- How many finished with no help.
- How many finished wrong with no warning, which is the dangerous kind.
- How long the repairs took.
Then decide what it can touch. If the second count isn’t zero, it doesn’t touch money.
Be clear about what 20 tasks can and can’t do. Each task is 5 points of your score, so one extra failure moves you more than the whole 2.6-point gap that made the headline. You can’t confirm or refute the benchmark with a sample that small, and you aren’t trying to. You’re looking for the kinds of failure your own work produces, and one example of a wrong figure filed without a flag tells you more than any average. Do a second round of 20 after a month, on fresh tasks, because vendors change models under the same name.
Sequence the permissions the same way. Start the model on read-only work, like pulling figures out of documents into a sheet someone else checks. Move to drafting. Leave anything that files, pays or sends for last, and only on tasks where the second count stayed at zero across both rounds.
A sceptic will say benchmarks like this are marketing and a 2.6-point lead is noise. Mostly right, and we wouldn’t base a purchase on the margin. But 47.3% to 75.0% in one release is hard to wave off, so something real changed in how well these models drive a desktop. We’d soften the caution if independent runs on messy business data landed near 75%. Until then, a trial is worth your afternoon, and nowhere near enough reason to skip it.
Enterprise and Edu access is listed as early access, so confirm that your plan includes it before you plan around it. Ask your reseller for a date in writing, since a plan that lists a feature as coming isn’t a plan that has it, and the test above needs the real thing, on your data, with your logins.
Frequently asked questions
Does GPT-5.4 beat humans at computer use?
On OpenAI’s reported test, yes by a small margin: 75.0% against a 72.4% human baseline. That baseline comes from the original OSWorld work and may not match your trained staff.
How much does GPT-5.4 cost through the API?
OpenAI lists $2.50 per million input tokens and $15 per million output tokens for standard use, against $1.75 and $14 for GPT-5.2. Requests beyond the 272K window count double against usage limits.
Should a small business let GPT-5.4 run desktop tasks?
No, not yet without a trial. Run 20 real tasks in a throwaway account, count silent errors, and only then decide what it can touch.
We have not tested GPT-5.4. Archive entry dated March 6, 2026, written and fact-checked on October 8, 2026. Sources are linked on the claims they support.
