AI math's pay-to-play era: open-source models can't keep up

A new benchmark called First Proof is putting a sharp number on a growing divide in AI: the models that can do research-level math are increasingly the ones with the biggest budgets. According to Scientific American, the benchmark confirms that some AIs are becoming exceptional tools for research-level mathematics, while open-source models still struggle. The result is a pay-to-play dynamic in which the best mathematical reasoning is concentrated in proprietary systems that cost serious money to build and run. For anyone building on AI, the message is blunt: if you want frontier math, you may need to pay for it.

First Proof is designed to test whether AI systems can handle the kind of mathematics that appears in research papers and advanced problem sets, not just textbook exercises. The name signals the bar: producing a first proof, the kind of rigorous logical argument that mathematicians spend careers learning to construct. The benchmark's findings suggest that the frontier of AI-assisted mathematics is moving quickly, but it is moving fastest for organizations that can pay for frontier compute and access. Open-source models, the kind that any developer can download and fine-tune, are not keeping pace on this particular measure.

The pay-to-play framing is not just about inference costs. Training and running state-of-the-art models requires massive GPU clusters, energy, and engineering talent. Open-source models, by contrast, are typically available for anyone to download and fine-tune, but they often lack the scale and specialized training that produce top-tier results on hard benchmarks. That gap has real consequences for startups, academic labs, and independent researchers who cannot afford to rent or build the infrastructure needed to compete. It also raises questions about who gets to participate in the next wave of mathematical discovery.

For the broader AI ecosystem, the benchmark adds evidence to a familiar debate: whether open-source AI can keep pace with closed, capital-intensive frontier models. In coding and general reasoning, open-source models have narrowed the gap. But in research-level math, the gap appears wider, possibly because mathematical reasoning demands long chains of logic and specialized training data that are expensive to curate and compute. The result is a two-tier system: well-funded labs get access to exceptional mathematical tools, while the open-source community is left working with models that still struggle on the same problems.

For executives and founders building on AI, the takeaway is practical. If your product depends on advanced mathematical reasoning, the cost structure of your AI stack just became a strategic variable. Teams that rely on open-source models may need to budget for proprietary APIs or invest in fine-tuning and infrastructure to close the gap. Academic institutions and public research labs, meanwhile, face a growing equity problem: the best tools for mathematical discovery may be reserved for those who can pay. That dynamic could shape everything from drug discovery to financial modeling to engineering simulation.

The First Proof benchmark is one snapshot, not a final verdict. Benchmarks can be gamed, and open-source models improve quickly. But the direction is clear: in AI math, access is becoming the differentiator. For anyone building the next generation of scientific tools, the question is no longer just which model is smartest, but who gets to use it. The answer, at least for now, is that the best mathematical reasoning is a pay-to-play sport.