The 2026 USA Mathematical Olympiad results are in, and the numbers are hard to overstate.

OpenAI’s GPT-5.4 (xhigh variant) scored 95.24% on the USAMO, a competition that requires six proof-based mathematical solutions over nine hours. Researchers at MathArena, an independent AI evaluation platform run by ETH Zurich, ran the benchmark shortly after the human competition concluded on March 22.

One year ago, top AI models scored near-zero on the same test.

What the Leaderboard Looks Like

MathArena used a jury of three frontier models to grade responses, specifically to counter the well-documented self-grading bias where models score their own outputs too generously. The results:

• GPT-5.4 (xhigh): 95.24%
• Gemini 3.1 Pro: 74.4%
• Claude Opus 4.6: 47%
• Step-3.5-Flash (open source): 44.6%

The gap between first and second place is wider than the gap between second and last. GPT-5.4 did not just beat the field. It essentially saturated the benchmark. Its one notable failure was Problem 5, where it incorrectly argued the statement was false and produced an invalid counterexample.

Claude Opus 4.6 also ran into a different kind of limit. The model exhausted its 128k-token context window on four of its 24 attempts, running out of space mid-proof. That is a computational ceiling, not a reasoning one, and one that will be addressed.

Why This Is Bigger Than a Math Test

The USAMO is not a trivia quiz or a multiple choice exam. Competitors work through six proof-based problems over two days. Solutions must be rigorous, original arguments, not pattern-matched answers. For decades, problems at this level required deep mathematical intuition that most humans never develop.

The fact that a commercial AI model now approaches near-perfect performance on it signals something specific: AI reasoning is no longer limited to surface-level pattern matching. These models are doing something closer to structured analytical thinking: forming arguments, checking internal consistency, and arriving at conclusions through logical chains rather than statistical associations.

For context: the top human competitors at the 2025 USAMO averaged around 70-80% over the same problem set. GPT-5.4 outperformed the human median by a significant margin.

What This Means for Business

The olympiad result is a useful signal because math has always been considered one of the clearest proxies for rigorous reasoning. It is harder to fake than language tasks, less dependent on memorised facts, and more dependent on thinking through novel problems.

If GPT-5.4 can do this on olympiad proofs, the practical upward revision of what AI can handle in business contexts is significant. A few implications worth paying attention to:

Analytical work is changing faster than organisations are adapting. Financial modelling, data analysis, strategic planning, legal reasoning: all of these rely on structured logical arguments. The tools available to professionals who use AI are now categorically different from what they were twelve months ago.

The capability gap between AI users and non-users is compounding. Professionals who have already built habits around working with frontier models are now accessing reasoning capabilities that were not possible in 2025. Those who have not are working with a progressively wider handicap.

The window for competitive AI adoption is narrowing. In early 2025, “we’ll get to AI soon” was a defensible position for many businesses. In mid-2026, that position is increasingly costly. If models can reason through mathematical olympiad problems, they can reason through your business problems: your pipeline analysis, your operational forecasting, your customer segmentation.

The Year-Over-Year Signal

The most important number from this benchmark is not 95.24%. It is the comparison to 2025, when top models scored near-zero on the same test.

That rate of improvement, from approximately 0% to 95% in twelve months on a benchmark considered one of the hardest quantitative challenges in existence, is the clearest evidence yet that AI capability is not plateauing. Whatever assumptions you made about AI timelines in 2024 or 2025 probably need revising upward.

For organisations still in evaluation mode, the time for that conversation is now, not next quarter.

---

If you want the playbook other teams are using with Claude and Codex right now, grab the free Working With Claude field guide. Download it here.