MATH-500: well-defined tasks
Cost rises 265x. Performance rises 2%. (r = 0.33*)
HLE: deep reasoning tasks
Cost rises 220x. Performance rises 5x. (r = 0.99)
All other models Mythos Preview Cost per review (top axis)
Cost = per full UW review (1.04M input + 110K output tokens). Mythos MATH-500 estimated at 99.5%. MATH-500 scores from Artificial Analysis independent benchmarks where available. HLE from published model/system cards. r = Pearson correlation between log-price and benchmark score. *MATH-500 correlation excludes estimated Mythos score; confirmed scores only.