GPT-5.4 Pro, OpenAI’s newest LLM, demonstrates an impressive long-term memory capability.
GPT’s mathematical prowess continues to advance, enabling it to tackle increasingly complex and challenging problems.
According to an January assessment by AI testing firm Epoch AI, GPT-5.2 Pro, an earlier iteration of the model, successfully resolved 31% of its mathematical tests, a significant improvement over its prior peak of 19%.
This latest version builds upon that momentum, successfully addressing a diverse array of mathematical challenges devised by academic experts.
An Epoch AI blog post indicates even greater strides: “GPT-5.4 Pro conquered a Tier 4 problem previously unsolved by any model. Initial analysis suggested it located a 2011 preprint, allowing it to circumvent significant portions of the task. The problem’s creator was not aware of this preprint,” the blog noted, clarifying that a preprint is an academic paper awaiting peer review.
Although it “solved” the problem, this highlights that GPT-5.4 – and all AI models – function as advanced information retrieval systems, with their efficacy tied to their ability to quickly access and process data.
GPT 5.4 Pro introduces various other advancements. OpenAI stated that this is the first iteration capable of actually performing actions on computers, instead of just describing how to do them. For instance, GPT-5.4 Pro possesses the ability to click a mouse, or more accurately, send a “click the mouse” instruction to an agent.
Furthermore, it boasts enhanced spreadsheet capabilities, efficiently resolves issues with fewer tokens, and generates a preliminary plan for intricate tasks, allowing users to fine-tune outcomes and guide its approach.
