AI gets about 10x cheaper every year (stevekrouse.com)

AI technology continues to experience rapid cost reductions, now confirmed to drop by approximately 10 times every year. This finding, originally posited by Sam Altman in February 2025, has been validated by new benchmarks across various assessments, including "Humanity's Last Exam" and "Arena Elo." For example, models like DeepSeek V4 Flash have showcased a steep price decline, from $1.93 per million tokens to just 18 cents in slightly over a year. However, the rate of decrease observed may slow down as earlier intelligence levels become too inexpensive for market interest, suggesting a limit on continual downward pricing. The significance of this trend for the AI/ML community cannot be overstated. As models become cheaper and more accessible, the potential for innovation and widespread integration of AI technologies grows exponentially, akin to the impact of Moore's Law on computing. In light of this, the proposed "Lee's Law" reflects this acceleration, indicating that inference costs are diminishing seven times faster than the transistor doubling rate established by Moore. This raises questions about the future market dynamics of AI as certain intelligence levels approach the threshold of being "too cheap to meter," and emphasizes the urgency for developers to capitalize on these advancements before they become commodities.