GPT-5.2 improves scientific and mathematical reasoning

OpenAI has introduced GPT-5.2 Pro and GPT-5.2 Thinking as its most capable models for supporting scientific and mathematical tasks. The company says both are better at following long chains of reasoning, keeping calculations consistent and working through complex technical problems.

The improvement is not limited to solving exercises. OpenAI argues that a model capable of reasoning with abstractions, staying consistent across many steps and applying what it has learned across different fields could be more useful for tasks such as programming, data analysis, experiment design and building simulations.

Results presented by OpenAI

In the GPQA Diamond test, a graduate-level exam covering scientific questions, GPT-5.2 Pro scored 93.2%. GPT-5.2 Thinking came close, with 92.4%.

In FrontierMath, an expert-level mathematics evaluation covering problems from levels 1 to 3, GPT-5.2 Thinking solved 40.3% of the exercises. OpenAI says this result sets a new record on the test.

These figures come from specific evaluations. They do not show that AI can do science without supervision. Performance can vary depending on the type of problem and the conditions of use.

How it can help in practice

In mathematics and theoretical areas of computer science, these models can help explore proofs, test hypotheses and identify connections between ideas. In other fields, they may be useful for preparing code, reviewing statistical analyses, organizing bibliographies or suggesting ways to investigate a question.

The important difference is the role the model plays. It does not replace the researcher or decide on its own whether a conclusion is correct. It can generate a detailed, well-structured argument, but it can also make mistakes, hide assumptions or present an invalid step as sound.

Researchers remain responsible for verifying accuracy, interpreting results and providing scientific context.

OpenAI reached this conclusion after working over the past year with scientists in mathematics, physics, biology and computer science. The company has also collected early examples from astronomy and materials science to show how GPT-5 was already being incorporated into real-world work. With GPT-5.2, it says these contributions are becoming more consistent and reliable.

What changes for you

If you work with technical information, the progress could mean a more useful assistant during the early stages of a project: understanding a problem, comparing approaches, writing a code draft or reviewing a chain of reasoning.

But verification remains essential, especially when medical, financial, industrial or scientific decisions are involved. A convincing result is not the same as a proven result.

The signal worth watching is whether these improvements hold up outside exams and lead to reproducible advances in real scientific projects. For now, GPT-5.2 points to a specific role: speeding up exploration and preliminary work while the responsibility for checking and understanding each result remains with humans.