Epoch AI analyzed API pricing and benchmark scores across six benchmarks over three years, finding that the cost to match a fixed capability milestone (e.g., GPT-4 performance on GPQA Diamond) has fallen at rates ranging from 9x to 900x per year, with a median of 50x. When only post-January 2024 data is used, the median rate rises to 200x per year, suggesting a recent structural acceleration rather than a smooth long-run trend. Declines are driven by a combination of smaller models, more cost-effective hardware, and potentially compressed margins, though the relative contribution of each is not formally modeled. For financial institutions, this rapid and accelerating cost deflation structurally compresses per-unit revenue for inference providers while simultaneously lowering barriers for application-layer entrants and enterprise adopters.
259 other tracked developments touching the same claims.