Benchmark price-performance falling 5–10× per year, but frontier model inference costs rising 3–18× per year as compute demands grow

A systematic study using the largest known dataset of historical AI benchmark prices (April 2024–November 2025, 93+ models across GPQA-Diamond, AIME, and SWE-bench) finds that the cost to achieve a given benchmark score is falling 5–10× per year for frontier models, driven roughly equally by algorithmic progress (~3× per year) and hardware/competition effects. Simultaneously, the cost to run frontier-level models is rising 3–18× per year, because marginal capability gains require disproportionately more inference compute, meaning roughly half of measured GPQA-Diamond progress disappears when inference cost is held constant. For financial institutions, this bifurcation matters structurally: cost-per-task economics favor AI adoption broadly, but the rising absolute cost of frontier inference constrains which use cases are economically viable and compresses the window before the next capability tier demands fresh spending.