GPT-5.4: the AI that now uses your computer better than you

On March 5, 2026, OpenAI launched GPT-5.4 and published a number that stopped the tech world for a few days: on the standard test measuring how well an AI can operate a computer, GPT-5.4 scored 75%. Expert humans score 72.4%.

For the first time, an AI surpasses the average human at operating a desktop computer. We're not talking about answering questions — we're talking about opening programs, navigating websites, searching for information, filling out forms, and completing real tasks on screen.

What can it actually do?

Imagine telling it: "Go to the 30 websites of these suppliers, find the price for this product on each one, and create a comparison table sorted from lowest to highest price." GPT-5.4 does it. It navigates each site, extracts the data, builds the table. While you do something else.

Before, automating that kind of task required hiring a programmer to write a custom script. Now you give the instruction in plain English.

What else does GPT-5.4 bring?

• 1-million-token context window: It can read and remember huge documents in a single conversation. A 500-page contract, a complete codebase, thousands of emails — all in memory at the same time.
• 33% fewer errors compared to the previous version. Responses are more accurate and reliable.
• Better with spreadsheets and presentations: Understands and works with Office documents more fluidly.

How big is the leap?

To put it in perspective: the previous version (GPT-5.2) scored 47% on that same computer-use test. In a single version, OpenAI went from 47% to 75%. It surpassed humans. That had never happened before with this capability.

The benchmark they used is called OSWorld and was specifically designed to measure whether an AI can complete real tasks on an operating system: opening apps, finding files, interacting with web pages, copying data between programs. It's the kind of work an administrative assistant does.

Is it available for everyone?

GPT-5.4's Computer Use is available primarily via API for developers. The ChatGPT version for regular users is incorporating these capabilities gradually. If you're a ChatGPT Plus or Pro user, you'll see it in the coming months.