AI gets about 10x cheaper every year An analysis published in October 2026 confirms Sam Altman's February 2025 claim that the cost of a given level of AI capability falls roughly 10x every 12 months, with the average across four benchmarks — Humanity's Last Exam, Arena Elo, GPQA Diamond and Terminal-Bench — showing a 10x price drop every 1.0 years. The per-benchmark rates ranged from 0.8 years (Humanity's Last Exam and Terminal-Bench) to 1.3 years (Arena Elo), with GPQA Diamond at 1.2 years. The author proposes calling the trend "Lee's Law," citing Kai-Fu Lee's May 2024 prediction of 10x annual inference cost declines, and notes the pace is 7x faster than Moore's Law's two-year transistor doubling. AI gets about 10x cheaper every year The cost to use a given level of AI falls about 10x every 12 months. – Sam Altman, February 2025 https://blog.samaltman.com/three-observations Now in October 2026, 20 months later, I wanted to see if that was still true, or if the rate had sped up or slowed down. 1 user-content-fn-making-charts Living on an exponential, it's really hard to get an intuition for it. It helps to zoom out and make charts. I charted the price decreases using four different benchmarks: | Benchmark | Time for 10× price drop | |---|---| | Humanity’s Last Exam | 0.8 years | | Arena Elo | 1.3 years | | GPQA Diamond | 1.2 years | | Terminal-Bench | 0.8 years | | Average | 1.0 years | I confirmed Sam's claim: AI continues to get about 10x cheaper every year. Humanity's Last Exam Humanity's Last Exam HLE tests AI models with difficult expert-written questions across a wide range of subjects. In this chart, each colored line represents a different level of intelligence on HLE. For each level of intelligence, you can walk along the graph to see subsequent models that would give you the same level of intelligence, or higher, but cheaper. For example, start with o3-mini, released in January 2025. It scored 12% on HLE and launched at $1.93 per million tokens. 2 user-content-fn-prices Just over a year later, DeepSeek V4 Flash offered that level of intelligence for about 18 cents per million tokens. That's 11x cheaper. But when you play it out further, the story isn't as clear. By September 2026, MiMo V2.6 Flash still cost about 18 cents per million tokens. That's only 11x cheaper in about 1.7 years, rather than the roughly 50x the rule would predict. I hypothesize that this flattening happens when things get too cheap to meter, and people don't care about making arbitrary, old levels of intelligence even cheaper. 3 user-content-fn-floor Averaging the five levels of intelligence gives 10x cheaper every 0.8 years . Arena Elo Elo measures which model's answers people prefer in head-to-head comparisons. It lets us look back further, to March 2023. Here the fitted rate is 10x cheaper every 1.3 years , which again might be slower because the oldest levels of intelligence aren't continuing to get cheaper. 3 user-content-fn-floor GPQA Diamond GPQA Diamond https://arxiv.org/abs/2311.12022 is a test of graduate-level science questions. The question here is the same: how much does it cost to get a model that scores at least 30%, 40%, 50%, and so on? 10x cheaper every 1.2 years . Terminal-Bench Terminal-Bench https://artificialanalysis.ai/methodology/intelligence-benchmarking measures how well models complete tasks in a terminal. 10x cheaper every 0.8 years . Lee's Law I admire Moore's Law and think we should have another law people can point to about how fast AI is getting too cheap to meter. I'd propose “Lee's Law.” The earliest prediction I found was Kai-Fu Lee predicting inference costs would get 10x cheaper a year in May 2024 https://www.jiemian.com/article/11166056.html .