AI enables greater access to software development and other knowledge-based activities. For $20/month or less you can access tools to speed up work, and work more easily in adjacent areas (and even sometimes non-adjacent: Claude helped me to fix my bathroom plumbing between coding sessions earlier this week).
However, this accessibility has limits in comparison to what we saw recently with OpenAI’s 88-hour sprint on their potential solution to the Navier-Stokes problem. This raises new questions: what are the consequences of the emerging capability gap between the AI labs and the rest of us in our ability to solve problems like this quickly, and what does that mean for the researchers to whom they are also suppliers?
It’s not that OpenAI or the big labs have a monopoly on solving maths problems with AI. Resolution of Erdős Problem #728: a writeup of Aristotle’s Lean proof and Primitive sets and von Mangoldt chains: Erdős Problem #1196 and beyond are both examples of public models being used to make leaps, and there are many more. The thing that’s unique here is the timescale the work was done on (88 hours), and the complexity of the problem targeted (this problem was one of seven Millennium Prize problems, specifically selected difficult, open problems).
We should recognise the independent progress on a precursor problem by two researchers – one of whom was an Anthropic employee – over a longer time frame. The impact of that progress on OpenAI’s own work is disputed. Assuming OpenAI’s work turns out to be valid, this was an impressive demonstration of their competitive edge.
For example, on the investment required: OpenAI consumed 130 billion output tokens in 88 hours to produce their proposed solution for Navier-Stokes. For an outsider, the output tokens alone would have cost ~$6.5MM at GPT-6 Astra API prices (some others claim a higher total cost). That outsider would also need the infrastructure and rate-limits to consume that many tokens in the time window. Neither the funding nor limits are typically available to researchers. Larger research programs can run into millions of dollars but these are spent on funding a program and its staff for years, not on a sprint of AI inference.
Of course, OpenAI didn’t pay retail API prices. They already had their own GPUs, infrastructure, and model. A scientist from the University of Michigan estimated their marginal cost for this work at “about $1 million in inference compute”. Further, the analysis was produced by scientists experienced with large-scale AI use and who had access to more capable unreleased models, infrastructure such as “a cached version of the Internet”, and a harness capable of orchestrating 10,000 agents.
Whilst it’s true that significant capital is already necessary in many disciplines (particle accelerators are hardly cheap), use of AI is a new need applying in disciplines where such investment was previously unexpected. Even as inference costs decrease over time and today’s unreleased models become public, the frontier will move forwards. As the creator of newer models and infrastructure, the labs obviously have early access. When demonstrating the superiority of these models they may find themselves disrupting or competing in fields beyond AI, with a unique advantage.
But how does this extend beyond maths? As a demonstration of the model’s power this work is compelling, but we should be careful what to infer from it. Maths proofs are verifiable within a closed system. Solutions to many problems in business and life cannot be verified with certainty before acting, which means 10,000 agents cannot be autonomously applied to them without questions of judgement and responsibility. Missiles cannot be unfired if targets were misidentified, and investment decisions cannot be undone if they lead to ruin. So whilst labs can still have an advantage, there’s a harder question about how to apply it in these other domains and therefore the gap is not as easy to take advantage of.
Still, this work illustrates the scale of the emerging capability gap in research. And, at some scale, it will spread to other fields over time if the market’s expectation of AI is real. In that future, more work needs to move beyond the modality of “chat assistant with maybe 5-10 agents” scientists and engineers outside the labs are familiar with today and towards broader automation, though this is still a work in progress.
So what can we do in the meantime? Making progress is worthwhile, even if we can’t individually close the gap fully. Experimenting with technologies and sharing learnings across your organisation can build expertise, e.g. a recent paper from Woodruff et al (2026) describes a number of general techniques along with examples of their use in solving real problems in research. Commercially-available tooling and open-weight models are also improving all the time, which might shift the economics of large-scale AI use. But gaps still persist, e.g. as an evaluation of Kimi K3 shows, in the security domain.
This doesn’t structurally alter the incentives of labs to continue disrupting other fields in the way that happened here. If competitive research continues trending towards requiring frontier-level AI capability, then access to frontier models and infrastructure at scale becomes a prerequisite, and access is largely controlled by a few large corporations today. Researchers are also grappling with the purpose this kind of demonstration serves, opening conversations like the open letter by a group of Fields Medalists arguing that creating understanding is as much an outcome as the solutions themselves.
Assuming the OpenAI proposed solution is valid, what structural change will take us from labs using AI to make breakthroughs in other fields, to that capability living within the fields themselves and, eventually, true democratisation of the technology?