OpenAI spent more than $1m of compute on a prize it says it will not claim

Start with the accounting, because it is unusually clean. Head of research Mark Chen says the compute bill exceeded the $1 million prize ^[4], the prize itself will not be claimed ^[6], and the Clay Institute's rules mean no claim was collectible for at least two years anyway: publication in a qualifying outlet, two years in public, broad acceptance by mathematicians ^[5]. So the $1m-plus was never spent chasing $1m. It was spent buying a demonstration, and Sebastien Bubeck's figure of roughly 10,000 agents at one point ^[7] tells you what the demonstration is meant to prove: that a fleet of agents can be pointed at an open problem and return something machine-checkable in Lean, whether or not a fluid actually blows up in finite time ^[3].

That is a marketing budget with a proof attached, and marketing budgets are the easy part. The line that will outlive the result is in OpenAI's own post: while unlikely, the company cannot rule out that de-identified data derived from Buckmaster and Alpoge's use of its products helped improve its models ^[9]. Call this a disclosure more than a defence. OpenAI's flat denial covers people and agents seeing the pair's prompts or drafts ^[8]; the carve-out covers the training pipeline, which is precisely the layer no user can audit and no lab can fully reconstruct. Alpoge, who works at Anthropic ^[10], credited them for saying it ^[11]. Buckmaster declined to allege anything, writing that he does not know whether their data was used ^[12].

Move from compute to credit: Buckmaster's statement says OpenAI floated proposals including removing Alpoge's name from the announcement ^[13], and Bubeck has denied ever asking for that, saying the screenshot was an attempt to coordinate releases ^[14]. Those two accounts cannot both be complete, and neither is checkable from outside.

The episode reads several ways at once. It could be boilerplate legal hygiene, the kind of language any lab with a usage-data training clause would write, and it matters only because the counterparties happen to be mathematicians with an unpublished result. It could be a genuine admission that de-identified telemetry from a paying user's session can flow into a model that then competes with that user, which is a live commercial exposure for anyone doing proprietary work in a hosted coding tool such as Codex ^[2]. Or it could be that the whole thing is about attribution etiquette between researchers, and gets settled by a footnote.

The second reading is the one that prices. If a research-heavy customer now has to assume that whatever cannot be ruled out in a blog post is a term of service, the cost of using a frontier lab's coding product on unpublished work rises without any price change on the invoice, and self-hosted or contractually ring-fenced alternatives get cheaper by comparison. What would prove this wrong: OpenAI publishing a data-lineage attestation specific to those Codex sessions, or the Clay Institute's eventual verdict ^[5] making the proof stand so clearly on its own construction that provenance stops being a question. Absent either, the exercise's durable output ends up being a sentence, with the theorem's status still unsettled.