GPT-5.4: The Game-Changing New Capabilities of OpenAI’s Latest Model – AI News – #2 March 2026

Just a few weeks ago, if you asked professionals about the best model for coding and analytical work, they often pointed to competing solutions like Claude. With the launch of GPT-5.4, OpenAI has effectively closed that gap.

What makes this update absolutely crucial is the convergence of advanced reasoning and coding into a single model. Instead of creating separate, highly specialized tools for different tasks, we are witnessing a shift toward unified models. GPT-5.4 can think through a problem from various angles: it will write the code itself, open a browser, gather data, and finally prepare a professional presentation from it. This direction of AI development is progressing much faster than most experts anticipated.

The most important indicator of the new model’s value is its performance on the GDPval benchmark. Why does this matter? Because unlike abstract benchmarks and flashy tech demos, GDPval measures the AI’s ability to generate actual, ready-to-deploy work deliverables across 44 different occupations.

[Placeholder: Image of an automated investment banking spreadsheet model]

The GPT-5.4 model set a new record here. In head-to-head matchups with human professionals, it matched or beat their results a staggering 83.0% of the time (for comparison, GPT-5.2 scored 70.9%). Its achievements in working with data look particularly impressive:

One of the most fascinating new features is equipping GPT-5.4 natively with computer use capabilities. The model perfectly handles mouse and keyboard operations based on screenshot analysis.

In the rigorous OSWorld-Verified test, which checks the ability to navigate a desktop interface, GPT-5.4 achieved a success rate of 75.0%. Interestingly, the average human score in the same test is 72.4%. This means that in standardized tasks involving clicking through applications, the model performs more efficiently than an average user.

OpenAI has also introduced a number of changes regarding the user experience:

Every revolution comes with a cost. To make GPT-5.4 more flexible and conversational, OpenAI engineers had to “loosen” its rigid boundaries slightly. As a result, the model gained a personality that some developers describe as “chaotic neutral.”

It sometimes overinterprets a task—it might implement a feature that no one asked for (e.g., adding GDPR consent checkboxes to a simple demo site), or “leak” part of the prompt directly into the UI elements. It’s akin to working with a brilliant, but occasionally rogue coworker.

GPT-5.4 is gradually rolling out across the OpenAI ecosystem. Currently, access to the “Thinking” variant and the “Pro” version (optimized for the most complex and compute-heavy tasks) is available to users on the ChatGPT Plus, Team, and Pro plans.

Regardless of minor hiccups, the convergence of analytics, programming, and understanding the visual world means that GPT-5.4 sets an entirely new standard in the AI industry.

Want to stay up to date with the latest changes in the AI and SEO world? The world of technology and search engines changes day by day. Subscribe to the Delante newsletter to regularly receive expert analyses, proven strategies, and the most important news that will help you stay ahead of the competition. Join the ranks of professionals today!

Source: https://openai.com/pl-PL/index/introducing-gpt-5-4/