The expense required for artificial intelligence to achieve a set accuracy level on five primary benchmarks has declined roughly 13x each year since 2023. Epoch found that the cost to hit a 75% score on the GPQA Diamond benchmark fell from $0.30 per question for OpenAI's o3 model to $0.0004 for GPT-5.6 Luna in less than 18 months, representing a 725x reduction. This study measures the total capital required to maintain a constant performance level over time rather than focusing on simple token pricing.

The cost for AI performance to scale is 54x faster than the decline of electricity costs up to 1973 and 18x faster than the drop in lithium battery prices, according to the research. The rates of decline also exceed those of compute and DNA sequencing by 6x and 4x, respectively. Epoch derived these trends by replaying benchmark transcripts under tighter token budgets using a methodology adapted from CAISI.