Epoch AI Data Records 47% Quarterly Drop in Cost per Fixed AI Performance Level

Epoch AI analysis by Emberson and Roodman documents a 47% average quarterly reduction in inference cost for constant benchmark performance over three years. This compounds to a 13-fold annual decline. The GPQA example illustrates a 725-fold reduction in under 18 months. Comparable historical rates for transistors or solar cells were slower and sustained over longer periods.

Data from the same report and cross-referenced with 2024-2025 MLPerf inference submissions show frontier models achieving equivalent accuracy at lower FLOPs per token. Open-weight releases lag frontier cost curves by 6-12 months even when parameter counts match. Lower inference prices directly expand viable use cases in real-time applications where token budgets previously constrained deployment.

Operational effects include expanded access to high-accuracy reasoning for non-enterprise users and reduced per-query margins for API providers. Daily workflows such as code assistance and document summarization become economically viable at higher volumes. The price trajectory compresses the window in which any single model architecture maintains a cost advantage.

Continued 40%+ quarterly declines imply that models meeting 2025 frontier thresholds will reach commodity pricing by late 2027, shifting value capture from inference hardware toward data curation and application integration.