MathArena employed a semi-automated, three-model judging panel (GPT-5.4, Gemini 3.1 Pro, Opus 4.6) to score solutions, aiming to reduce self-bias and formatting bias, with human-review score shifts capped at two points.

The evaluation incurs notable costs: GPT-5.4 at $5.15 per run, Gemini 3.1 Pro at $2.20, Claude Opus 4.6 at $13.23, and Step-3.5-Flash at $0.22.

GPT-5.4 (xhigh) led the field with a 95.24% score on the 2026 US Mathematical Olympiad (USAMO).

GPT-5.4’s sole significant error occurred on Problem 5, where it produced an invalid counterexample and incorrectly argued the statement was false.

From 2025 to 2026, improvements include not only higher scores but also a shift in error types toward fewer circular arguments and incoherent proofs; remaining mistakes are subtler, such as occasional mid-proof chain-of-thought remnants.

The results indicate frontier AI is nearing expert-level mathematical reasoning, underscoring the pace of AI-driven breakthroughs in science.

Gemini 3.1 Pro finished second with 74.4%, followed by Claude Opus 4.6 at 47% and open-source Step-3.5-Flash at 44.6%.