AI Inference Cost per Performance Level Falls 47% Quarterly, Epoch AI Report Finds

The study highlights a dramatic shift in the economics of reasoning models. For example, attaining a 75% score on the GPQA Diamond benchmark cost approximately $0.30 per question with OpenAI's o3 in January 2025. By mid-2026, OpenAI's GPT-5.6 Luna achieved a similar score for roughly $0.0004 per question—a 725-fold reduction in cost in under 18 months.

The rate of decline varies depending on the task. Mathematical problems saw the fastest progress at 50–52% per quarter, while game-based puzzles declined by roughly 39–43% per quarter. On average, the decline rate is highest near state-of-the-art (SOTA) performance levels, dropping from an average of 66% per quarter at the moment of SOTA debut to 32% per quarter two years later.

This rate of decline is significantly faster than other transformative technologies. The observed drop in AI costs is four times faster than the decline in DNA sequencing and 54 times faster than the price drop of electricity. The report notes that price competition appears particularly intense among frontier models.

PLUS ULTRAby Amenoyomi

This decline is measured not by the price per token, but by the cost required to reach a specific benchmark score. This distinction is critical for reasoning models, which often consume significantly more tokens to achieve high performance but may utilize an underlying model that is smaller and cheaper to run per token.

The reduction in cost follows a specific cycle tied to the state-of-the-art (SOTA) frontier. When a new performance level is first achieved, providers can briefly charge a price premium. However, as competitors catch up and technological improvements accelerate, prices plunge rapidly—averaging a 66% quarterly drop at the moment of SOTA debut, which slows to approximately 32% per quarter after two years.