Rohan Paul
@rohanpaul_ai
AI costs are indeed dropping dramatically.
On GPQA Diamond, Epoch estimates OpenAI's o3 model needed about $0.30 per question to reach a 75% score.
Under 18 months later, GPT-5.6 Luna matched that level for $0.0004, a 725x cost drop.
Overall, Epoch estimates matching a fixed benchmark score has become about 47% cheaper each quarter since 2023.
That works out to roughly a 13x annual cost reduction across its five primary benchmarks.
Unlike earlier token-price studies, Epoch measures the dollars needed to reach the same benchmark accuracy over time.
It estimates each model’s cost-performance curve by replaying benchmark transcripts under tighter token budgets, using a method adapted from CAISI.
AI’s decline is 54x faster than electricity, 18x than batteries, 6x than compute and 4x than DNA sequencing.
