I ran GPT-6 Luna Max on MathArena's harness

A developer tested GPT-6 Luna Max on MathArena's harness, noting its strong performance on Riemann Bench and considering it an underrated model. The model correctly answered 16 out of 19 finite or discrete questions, 14 out of 20 analysis and probability questions, and 9 out of 18 geometry, algebra, and topology questions. The developer abandoned testing another model, xhigh, after it performed worse than Luna Max.

This report uniquely offers specific performance metrics for GPT-6 Luna Max across different mathematical domains, unlike general claims of strong performance.