Both sides, in good faith

The strongest fair case each way — we don't pick a winner.

The case for

Buckmaster and Alpöge have reasonable grounds for concern: their work was developed openly using Codex and Claude, OpenAI's own account leaves open the possibility that de-identified usage data shaped its models, and the timing—an announcement one day after their publication—is striking. In a field that depends on trust and proper attribution, researchers are entitled to expect straight answers about whether their unpublished work informed a rival's output, and OpenAI's reportedly evasive responses understandably deepen suspicion rather than allay it. Given the reputational and financial stakes of claiming a Millennium Prize-calibre result, it is not unreasonable to demand full transparency about training data and timelines before accepting the achievement at face value.

The case against

OpenAI and Sebastien Bubeck have offered a coherent account: a more powerful, unreleased internal model working with thousands of concurrent agents began training in late August, independent of and prior to seeing the rival team's published work, and the two proofs are said to differ substantially in approach. It is entirely plausible that two well-resourced teams working on closely related mathematical problems, using similar AI-assisted tooling, would converge on breakthroughs within days of each other by coincidence rather than misconduct. Acknowledging an inability to rule out every theoretical trace of de-identified data in a model trained on vast, aggregated usage logs is a reasonable statement of epistemic humility about how large models work, not an admission of impropriety, and OpenAI's disavowal of the prize money suggests it is not seeking to profit unfairly from the result.