GPT just keeps getting better at mathematics, increasingly solving the trickiest of problems.

In January, AI testing company Epoch AI found that a previous version of the AI model, GPT-5.2 Pro had solved 31% of its mathematical challenges, a rise from the previous score best of 19%.

The newest version has continued the progress, tackling a range of mathematical problems set by academics..

According to an Epoch AI blog post, it has gone one stage further: “GPT-5.4 Pro solved one Tier 4 problem that no model had solved before. In a preliminary analysis, it appeared to have found a preprint from 2011 which let it shortcut much of the intended work. The problem author was unaware of this preprint,” said the blog, referring to an academic paper that had not yet been subjected to peer review.

While “solving” the problem, this does demonstrate how GPT-5.4 (and indeed all AI models) are sophisticated search engines, whose effectiveness depends on access to and rapid assimilation of information.

GPT 5.4 Pro offers several other innovations. Open AI claimed that it was the first version that can do things on computers, rather than saying how to do things. For example, GPT-5.4 Pro can click a mouse — or to be more precise, it can issue a “click the mouse” command to an agent.

It also has improved spreadsheet skills, can solve problems using fewer tokens, and will prepare a plan before tackling complex tasks, giving users the chance to tweak the results and steer it in the correct direction.