OpenAI Tests Model on 4,000 Unsolved Math Problems

For years, math tests have served as the standard way to evaluate AI models such as ChatGPT and Claude. OpenAI now claims its models have become so proficient that traditional math tests no longer challenge them enough. That is why researchers turned to a more difficult approach.

An unreleased OpenAI model was given approximately 4,000 unsolved math problems. The outcomes were both surprising and troubling.

From this exercise, the AI generated 722 manuscripts, which were organized into 372 families of related findings. According to OpenAI, each result consumed an average of roughly three hours of reasoning computation.

“Some pretty exciting days ahead for the mathematical community!” said Stefano Gogioso, a member of BeInCrypto’s Future Tech and AI Experts Council

Is AI Advancing at an Unsafe Pace?

The bigger story is that frontier AI is starting to go beyond answering known questions and into proposing possibilities for unknown ones.

This development could greatly expand the amount of intellectual work a researcher, an engineer, or an analyst can take on.

However, output does not equal truth. Several of OpenAI's papers include computer-checkable Lean proofs, while others lack them. OpenAI cautions that some unverified results "could have issues."

This introduces a fresh bottleneck: human researchers may find it difficult to verify findings as fast as AI can generate them.

We’re releasing a broad range of new mathematical results produced by an internal frontier model.

We’ve been consulting with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study, and we have drawn on their advice and…

— OpenAI (@OpenAI) October 6, 2026

Could an AI Conduct Predictive Analysis for Investing?

It is possible, but finance presents a distinct set of challenges.

A mathematical proof can ultimately be proven correct or incorrect. Markets, however, are noisy and perpetually in flux.

Recent benchmarks in finance indicate that frontier AI still faces difficulties with complex investment research, and studies on market timing reveal only limited predictive benefits.

In the near term, the opportunity lies in deeper analysis, not flawless prediction.

An artificial intelligence that can reason for hours on end might simultaneously scrutinize filings, earnings calls, macroeconomic data, and competing scenarios, subsequently testing many more hypotheses than a single analyst could.

That might be the more significant takeaway from OpenAI's test. AI is increasingly able to generate serious analytical output at an enormous scale. The next challenge is determining which of that output is worthy of trust.