GPT-5.4 Is Here, OpenAI’s First Model That Can Actually Use Your Computer
On March 5, 2026, OpenAI released GPT-5.4 — and for the first time, a general-purpose AI model from OpenAI can natively control a computer. It can move a mouse, type on a keyboard, navigate between applications, fill in spreadsheets, and complete multi-step workflows across software environments entirely on its own.
That single capability change makes GPT-5.4 the most consequential OpenAI release since GPT-4. Here’s everything you actually need to know.
What GPT-5.4 Is
GPT-5.4 is OpenAI’s new flagship foundation model for professional work. It ships in three variants:
- GPT-5.4 Thinking — the standard version, available to all paid ChatGPT subscribers (Plus, Team, Pro)
- GPT-5.4 Pro — higher-performance version for ChatGPT Pro ($200/mo) and Enterprise customers
- GPT-5.4 via API & Codex — for developers; supports up to 1 million tokens of context
GPT-5.2 Thinking will remain available as a legacy option until June 5, 2026, giving users time to migrate.
The Computer Use Capability — And Why It’s A Big Deal
GPT-5.4 is the first general-purpose OpenAI model with native computer-use capabilities. It reads screenshots, controls mouse and keyboard commands, navigates browsers, edits documents, and manages workflows across multiple applications — without any developer needing to build that infrastructure themselves.
On OSWorld-Verified — the industry benchmark that measures how well an AI can navigate desktop software — GPT-5.4 scored 75.0%. Human performance on the same benchmark is 72.4%. That means GPT-5.4 is now statistically better than the average human at using a computer in test conditions.
“GPT-5.4 can execute mouse and keyboard commands based on screen screenshots, automating complex workflows across software and web environments.” — Digital Today
Benchmark Results
| Benchmark | GPT-5.4 | GPT-5.2 | What It Tests |
|---|---|---|---|
| OSWorld-Verified | 75.0% | 47.3% | Desktop computer use (human avg: 72.4%) |
| WebArena-Verified | 67.3% | 65.4% | Browser automation |
| GDPval | 83.0% | 70.9% | Real professional tasks across 44 jobs |
| SWE-Bench Pro | 57.7% | ~40% | Real-world software engineering |
| Spreadsheet modelling | 87.3% | 68.4% | Finance & enterprise data work |
| False claim rate | 33% lower | baseline | Hallucination reduction |
The Efficiency Story
Beyond raw capability, GPT-5.4 is significantly more token-efficient than its predecessors. OpenAI reports it uses up to 47% fewer tokens on some tasks than GPT-5.2, which means, despite being priced slightly higher per token, it can actually cost less in practice by doing more with fewer tokens.
In a real-world test by Mainstay, GPT-5.4 completed sessions across ~30,000 property portals with a 95% first-attempt success rate, running three times faster and using 70% fewer tokens versus prior computer-use models.
New Enterprise Features
OpenAI also announced direct ChatGPT integrations with Microsoft Excel and Google Sheets alongside the GPT-5.4 launch. The model can be plugged directly into spreadsheet cells, enabling granular analysis and automated task completion inside the tools millions of knowledge workers already use daily.
A new “Tool Search” system in the API lets agents dynamically find the right tool from a large ecosystem, cutting token usage by up to 47% in multi-tool workflows. GitHub’s Chief Product Officer, Mario Rodriguez, said: “Developers don’t just need a model that writes code. They need one that thinks through problems the way they do.”