← sources

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

source: reddit (reddit:singularity)published 2026-09-26 19:52ingested 2026-09-27 02:38

retained text

I still find GPT-6 Luna Max to be one of the more underrated models. It also did quite well on Riemann Bench. I was originally testing xhigh, but as soon as it got two more wrong than Max, I decided to abandon it. 

I forgot that MathArena does theirs on batch, so you would probably see roughly 40-50% decrease in cost per problem. 

If anyone is interested in seeing the results, I can put it in a repo. Otherwise, it matched 16/19 finite or discrete answers, 14/20 analysis and probability answers, and 9/18 geometry, algebra, and topology answers. 

document fetched 2026-09-27 00:30 · canonical https://reddit.com/r/singularity/comments/1wqzvgo/i_ran_gpt6_luna_max_on_matharenas_harness