OpenAI's GPT-5.2 Pro has added another notable result to the fast-moving debate over AI and mathematics: it helped solve Erdős problem #281 from number theory.
The result matters because mathematician Terence Tao described it as "perhaps the most unambiguous instance" of an AI solving an open mathematical problem. But the same story also carries a warning. A new database tracking AI attempts at Erdős problems shows that success remains rare, with actual success rates of just one to two percent.
What GPT-5.2 Pro Helped Solve
Neel Somani used GPT-5.2 Pro to crack Erdős problem #281, a problem from number theory. The source describes this as another Erdős problem that OpenAI's model has helped solve.
The significance is not only that an AI model was involved, but that Tao viewed the case as unusually clear. His phrase, "perhaps the most unambiguous instance," places this example in a different category from results where the role of the model may be harder to judge.
There is still an important caveat. Earlier proofs may have influenced the model's answer. That matters because, in mathematical work, the distinction between reproducing, recombining and genuinely contributing can be difficult to assess.
Even with that caveat, Tao confirms that GPT-5.2 Pro's proof is "rather different." That detail is central to why the case has drawn attention: the model's output was not presented as merely the same route in slightly altered language.
Why Tao Urges Caution
Tao also warns against a skewed perception of what AI can currently do in mathematics. The public picture can become distorted because negative results rarely get published, while positive results go viral.
That imbalance is easy to understand. A successful proof attempt is clear, memorable and shareable. A failed attempt is less visible, even though failed attempts are essential for judging the actual capability of a system.
For readers following AI math progress, this distinction is crucial. One solved Erdős problem can show that a model is useful in a serious mathematical setting. It does not, by itself, show that the model can reliably solve open problems across the board.
The source presents Tao's view as a balance between usefulness and limits. AI serves as a useful tool here, but moderately difficult Erdős problems might remain out of reach, according to Tao.
The New Database Changes the Picture
A new database by Paata Ivanisvili and Mehmet Mars Seven tracks AI attempts at Erdős problems. Its findings put the GPT-5.2 Pro result into a wider frame.
The database shows actual success rates of just one to two percent. The successes are also clustered around easier problems.
That context matters because it shifts the question from whether AI can ever help solve an open mathematical problem to how often it can do so. The answer, based on the database described in the source, is that most attempts still fail.
The database also helps explain why isolated wins can be misleading when viewed alone. A single success may be real and important, but it can coexist with a large number of unsuccessful efforts that receive far less attention.
How to Read the Result
The GPT-5.2 Pro result should be read as evidence of real progress, not as proof that AI has broadly mastered open mathematical problems. It shows that an AI model can be a meaningful part of a successful effort on an Erdős problem. It also shows why expert verification remains central.
Several facts sit side by side:
- OpenAI's GPT-5.2 Pro helped solve Erdős problem #281.
- Neel Somani used the model in the work.
- Terence Tao called the case "perhaps the most unambiguous instance" of an AI solving an open mathematical problem.
- Tao confirmed the proof is "rather different" despite the possibility that earlier proofs may have influenced the answer.
- A database of AI attempts shows success rates of just one to two percent.
Taken together, those points suggest a more grounded view of AI in mathematics. The technology can produce results worth serious attention. At the same time, its failures remain the dominant pattern in the tracked attempts described by the database.
A Useful Tool, Not a Guaranteed Solver
The broader implication is that AI's role in mathematics may be most productive when treated as a tool rather than a replacement for mathematical judgment. The source does not present GPT-5.2 Pro as independently resolving every challenge. It presents a case where the model helped in a result that experts found noteworthy.
The first autonomous solution to an Erdős problem confirmed by Tao dates back to January 4, 2026. The new GPT-5.2 Pro case extends that story, but the database adds a necessary counterweight: most attempts still do not succeed.
For now, the clearest takeaway is measured. AI models are becoming relevant to open mathematical work, including Erdős problems. But the available evidence in the source points to a field where successful cases are still rare, especially beyond easier problems.