On 6 October 2026 OpenAI published 722 mathematical manuscripts produced by an internal model. Artificial intelligence has just handed us a collection of research results, organised into 372 families and at different stages of verification, freely available at this link.
This is HUGE news, and it forces an uncomfortable question on us: how long will it take us to understand what we have just received?
We are used to asking machines to speed up our work. In this case, however, it is the machine asking us to pick up the pace: the paradigm has just been turned on its head. We have been overtaken.
What OpenAI’s mathematical manuscripts contain
ChatGPT can never have sung Venditti’s famous line “maths will never be my trade”, because after solving the Millennium Problem on the Navier-Stokes equations in September, it has just revolutionised mathematics by publishing 722 manuscripts.
Each family groups related work: a main result, its consequences or alternative proofs. OpenAI warns that some non-formalised results may contain errors. It will take time, clearly: yet this is still an extraordinary result, one that revolutionises the way research is done and opens up scenarios impossible to imagine just a year ago. Now verifying it all is up to us.
But how? Mathematical manuscripts need capable readers. And the expertise required for such work depends on the problem tackled. Since we are talking about problems that had never been solved before, finding errors in an advanced proof will require a degree of specific training.
No shortcuts, in short.
Three hours of compute, and the time it takes to understand
OpenAI states that the average result consumed compute equivalent to roughly three hours of ChatGPT Pro thinking. The company has also published ten summaries of the model’s reasoning.
Very little, on closer inspection: but there is a very evident asymmetry worth thinking about. Solving a problem took the machine just 180 minutes of extreme computation, yet verifying those calculations will take a researcher days, or weeks.
This opens up an interesting scenario, one not yet assessed to date: artificial intelligence could make the capacity to verify an increasingly scarce resource relative to the volume of material produced.
It is a hypothesis suggested by the very scale of this experiment, which the entire academic community is mobilising to verify.
Mathematicians call for more than publication
The Advisory Group on Mathematics and Artificial Intelligence, the independent group consulted by OpenAI, made this clear in its note of 6 October: its advisory role constitutes neither a judgement on the impact of the results nor an endorsement of the process followed by the company. The assessment rests with the mathematical community.
The group regards publication as the start of the process of understanding the work and incorporating it into mathematical knowledge. It also stresses that researchers must be able to formulate their own questions, rather than merely studying results selected by AI laboratories.
In its recommendations of 29 September, AGMAI calls on laboratories to support human understanding of their outputs financially as well, and to prevent mathematical results from becoming marketing vehicles. It also criticises the use of proprietary models inaccessible to the community to tackle advanced problems.
In response, OpenAI is announcing funding for workshops, conferences and programmes dedicated to understanding the main results produced by AI.
The value of this commitment will depend on how it is delivered. Publishing the mathematical manuscripts makes it possible to examine them. Enabling people to understand them requires resources, access and research freedom.
What this experiment tells us about superintelligence
If until just a few days ago superintelligence seemed (Trump’s words notwithstanding) a distant milestone, today we cannot be quite so sure. Granted, a collection of mathematical results does not prove that human capabilities have been surpassed in every field. Without doubt, however, academia has never been under so much pressure. Both in terms of verification and in terms of comprehension.
The question is this: if a system can propose answers to complex problems, how well equipped are we to recognise the valid ones, reconstruct their meaning and use them without relying solely on the authority of whoever published them?
It is a question that also runs through the journey described in our article on Claude Science, which launched this strand of AI applied to science. The common thread is the possibility of expanding scientific work without losing the ability to judge its results.
Here OpenAI presents us with a collection to be assessed. The promise of accelerating research meets the work needed to turn that acceleration into shared knowledge.
For those running a company, the question is already useful
Mathematics or not, the problem remains. At this point, every CEO should ask what happens when the capacity to produce outstrips the capacity to control.
Imagine a team able to prepare far more analyses thanks to artificial intelligence. Before counting that number as a productivity gain, it is necessary to understand how many analyses can be verified and how many genuinely improve a decision.
If the time saved in production is absorbed by review, the benefit changes. If review is compressed to keep up with the new pace, the likelihood of an error going unnoticed increases.
This is where Human Oversight comes in: assigning control to a person makes sense when that person has the expertise and the time to exercise it.
That is why, when facing an investment in AI, a useful question is: what verification capability do we need to build alongside the generation capability?
This may mean making sources accessible, introducing appropriate automated checks, defining who approves outputs and choosing which activities merit in-depth review. The answer depends on the work and on the consequences of error.
The mathematical manuscripts published by OpenAI make this tension visible on an unusual scale. It would be premature to claim that machines already produce knowledge faster than humans can absorb it.
Yet it is clear that all of this will happen sooner or later.
To genuinely accelerate the production of answers, we should first understand how to grow our capacity to judge them.
Domande frequenti
How many mathematical problems OpenAI has solved
The figure refers to the documents published. The families group related results, and verification is at different stages. It should not be read as a count of independent, validated solutions. [1]
How proofs are verified in Lean
A complete proof accepted by the checker is verified against the statement, the axioms and the dependencies used. It must also be confirmed that the formalised statement matches the original problem. Formal checking does not, on its own, establish the scientific importance of the result. [3]
What this announcement shows about superintelligence
The announcement does not demonstrate a general capability superior to that of humans. That would require evidence across many different domains. Instead, the experiment raises a specific question about the relationship between automated production, verification and understanding in research.