Back

OpenAI just released GPT-5.4: here’s what you need to know

Deveshi Dabbawala

March 6, 2026
Table of contents

The new model can use your computer, work through a million words at once, and handle tasks that usually need a professional.  

And that model is OpenAI’s GPT 5.4.

About GPT-5.4

OpenAI released GPT-5.4 on March 5, 2026. GPT-5.4 is available in two forms, namely GPT-5.4 Thinking, which is for everyday use, and GPT-5.4 Pro, which is for heavy-duty use. GPT-5.4 is available via ChatGPT and the OpenAI API.  

What’s new with this version?  

So far, OpenAI’s versions of AI models each had their specialty. Some were good at coding, and some were good at step-by-step thinking. GPT-5.4 does all of this and more in one go. Now, you won’t need to use other tools to get the best results, as GPT-5.4 does it all for you.  

GPT-5.4 is also more efficient, using up to 47% fewer tokens for certain tasks compared to  

other versions of the model. For developers, this means lower costs for using AI at scale.  

83% 

Matches professional across 44 occupations 

75% 

OSWorld desktop navigation (beats humans) 

 

33% 

Fewer hallucinations vs GPT-5.2 

1M 

Token context window in API & Codex 

 

What can it actually do?

GPT-5.4 moves beyond answering questions. It can perform real tasks, work with complex data, and interact with software tools.

Use your computer

GPT-5.4 can operate desktop apps, click menus, fill forms, and complete tasks on its own. It solved 75% of desktop tasks in tests, slightly higher than the human score of 72.4%.

Handle massive documents

The model can process up to one million tokens at once. That equals hundreds of research papers, full contract libraries, or large codebases in a single session.

Make fewer mistakes

Individual claims are 33% less likely to be incorrect than GPT-5.2. Overall responses are 18 percent less likely to contain errors.

Work inside Excel and Google Sheets

Plugins allow the model to connect directly to spreadsheets, update cells, run calculations, and complete tasks without manual copy and paste.

Select the right tools automatically

The new Tool Search engine lets the model identify and use the tools it needs without developers listing them in every prompt.

The data shows the reality

OpenAI's benchmark results make a strong case. On OSWorld-Verified the gold standard for desktop computer-use tasks GPT-5.4 hits a 75% success rate, a dramatic leap from GPT-5.2's 47.3% and notably above the 72.4% baseline achieved by human testers.

Benchmark 

GPT-5.2 

GPT-5.4 

Human / SOTA 

OSWorld-Verified (desktop navigation) 

47.3% 

75.0% 

72.4% (human) 

BrowseComp (deep web research) 

 

+17% absolute 

Pro: 89.3% SOTA 

GDPval (knowledge work, 44 occupations) 

 

83% 

Matches/exceeds professionals 

MMMU-Pro (multimodal vision) 

79.5% 

81.2% 

 

WebArena-Verified (browser tasks) 

65.4% 

67.3% 

 

The GDPval result is perhaps the most striking: in a test spanning sales presentations, accounting spreadsheets, urgent care schedules, and manufacturing diagrams, GPT-5.4 matched or exceeded industry professionals in 83% of comparisons across 44 occupations.

Why does this matter for businesses?

OpenAI is targeting enterprise users with GPT 5.4. Its advantages are that it can execute long multi-step tasks, run software directly, and integrate with tools that the company is already using. The concept is simple: instead of telling the AI what to do, it does the work itself.

For developers, it means reduced costs due to fewer tokens used. For business teams, the advantage of using the GPT 5.4 Thinking model is that it shows the plan before execution and the reasoning behind it, making AI execution simpler.  

“The result is a model that gets complex real work done accurately, effectively and efficiently delivering what you asked for with less back and forth.”

— OpenAI

GoML's take

Monitoring is in the progress of frontier model development at GoML, and GPT-5.4 is the type of product that makes the whole industry adjust its pace and perspective. Bringing coding, reasoning, and computer use together in one efficient model is both a technical breakthrough and a smart strategic move.

OpenAI is making a bet on the future of AI value, and it is no longer in specialized models for specialized functions, but rather in a single model that can handle the full scope of professional functions.  

Whether or not this will redefine the way in which the enterprise market approaches AI, further propel the agentic AI movement, or simply raise the bar for what is possible with a foundation model, one thing is for certain the age of AI that only reads and writes is dead and gone.  

GoML is here to help you take your AI projects and turn them into operational solutions with AI Matic, our framework for building and scaling enterprise AI solutions.