Source: Epoch AI | Published: 2026-02-22

Summary

Epoch AI reports that Google’s Gemini 3.1 Pro performs comparably to its predecessor Gemini 3 Pro on the FrontierMath benchmark — a collection of original, research-level mathematics problems designed to be beyond current AI capabilities. The headline result is incremental: no significant leap between model generations. However, the buried lede is more interesting: during a second evaluation run, Gemini 3.1 Pro solved a Tier 4 problem (the highest difficulty level) that no AI model had previously cracked. The problem was created by mathematician Emmanuel Breuillard. The caveat: the solution approach was described as “not how a human would” solve it, suggesting brute-force or unconventional methods rather than mathematical insight.

Key Insights

The stagnation on FrontierMath between Gemini 3 Pro and 3.1 Pro is notable because it contrasts with the rapid progress seen on easier benchmarks. FrontierMath was specifically designed to resist the kind of pattern-matching that drives performance on standard math benchmarks, and it appears to be working — model generations are not translating into proportional gains here. The Tier 4 solve is interesting precisely because of its non-human character: it suggests that AI math capabilities may advance through alien methods rather than by learning to reason the way mathematicians do. Whether this counts as “progress in mathematics” or “progress in computation” is a question the field hasn’t resolved. Epoch notes that evaluation of Gemini 3 Deep Think (Google’s reasoning model) was pending API access, which could be the more consequential result when it arrives.

Our Take

This is a useful data point in the ongoing debate about whether scaling is producing genuine mathematical reasoning or increasingly sophisticated pattern completion. The fact that a Tier 4 problem was solved at all is significant — these are problems designed by active researchers to be hard for AI — but the “not like a human” qualifier is doing a lot of work. For prediction market purposes, this reinforces the view that frontier math benchmarks will remain stubbornly resistant to rapid improvement, which is relevant for markets on AI capabilities milestones. The pending Deep Think evaluation is what to watch — reasoning models have shown disproportionate gains on hard math problems, and Google’s entry into that space could shift the picture.