OpenAI launches GPT-4o-Math to improve relations with mathematicians

OpenAI announced on Monday that it will roll out a new version of its ChatGPT model, dubbed GPT‑4o‑Math, aimed at tackling advanced mathematical problems, after a series of public missteps that have strained its relationship with the research community. The company said the upgrade will be available to all ChatGPT Plus subscribers starting next week and will be showcased at a virtual symposium with the Clay Mathematics Institute on Thursday. The move follows criticism that OpenAI’s earlier claims of “breakthroughs” in mathematics were either overstated or poorly communicated.

The controversy matters because AI tools are increasingly being used to explore conjectures, verify proofs and generate new ideas in pure math. In early 2024, OpenAI’s GPT‑4 demonstrated strong performance on benchmark tests such as the MATH dataset, prompting excitement that language models could become research assistants. However, the excitement turned to skepticism when the firm announced a supposed proof of a long‑standing combinatorial conjecture that later proved to be an error in the model’s reasoning. The incident highlighted the difficulty of trusting unverified AI output in a field where rigor is essential.

Mathematicians have pointed to a pattern of rushed announcements and insufficient peer review. In March, a blog post from OpenAI claimed its model had found a novel approach to a problem in algebraic topology, only to retract the claim after independent experts could not reproduce the result. The Institute for Advanced Study’s mathematics department issued a statement warning that “AI‑generated proofs must be treated with the same scrutiny as any human work.” OpenAI’s chief scientist, Mira Murati, acknowledged the missteps and said the company will involve external reviewers before publicizing future results.

OpenAI says the new GPT‑4o‑Math will undergo a formal validation process involving a panel of mathematicians from universities and research institutes. The upcoming symposium will feature live demonstrations, a Q&A with the panel, and a pledge to publish all model‑generated proofs in open‑access repositories for community review. If the vetting process works, the upgrade could restore confidence and give researchers a reliable tool for exploring complex problems. The next few months will test whether OpenAI can balance rapid innovation with the careful standards that mathematics demands.