AI Performance Costs Fall Faster Than Any Previous Technology
Epoch AI measures a 13-fold annual decline in the cost of reaching fixed AI benchmark performance, while MIT estimates algorithmic progress alone at roughly 3x per year. Reasoning models can still be expensive, and quality, speed, and error rates matter as much as price.
The cost of reaching a fixed level of artificial intelligence performance is falling faster than for any previous technology, according to new analyses that separate raw price declines from genuine algorithmic progress. Epoch AI, a research organization that tracks trends in machine learning, measures a price decline of about 13 times per year for AI systems hitting a given benchmark score. That figure captures the combined effect of better hardware, more efficient software, and intensifying competition across the industry.
Stripping out those hardware gains and market pressures, researchers at MIT estimate that algorithmic progress alone accounts for roughly a threefold annual improvement. In other words, even without faster chips or cheaper electricity, the same AI capability would become about three times less expensive each year because of smarter model design, better training methods, and more efficient architectures. That pace of algorithmic improvement is itself unusual when compared with historical technology curves.
The finding does not mean that the most advanced AI models available today are cheaper to use. Reasoning models, which spend additional computation working through problems step by step before producing an answer, can cost significantly more per task because they consume far more processing power. The headline price decline applies to reaching a fixed performance level, not to the absolute cost of the most capable systems on the market.
For organizations and individuals choosing an AI model for practical use, price is only one factor among several. Quality of output, response speed, and error rate matter just as much, and in many applications they matter more. A model that is cheap but unreliable may end up costing more in human review and correction than a pricier system that gets things right the first time.
The rapid decline in cost per unit of performance has broad implications. It suggests that capabilities once available only to well-funded laboratories and large corporations are becoming accessible to smaller businesses, independent developers, and public institutions. Tasks that were economically unviable a year ago may now be affordable, and tasks that are unaffordable today may become routine within months.
At the same time, the distinction between benchmark performance and real-world usefulness remains important. A fixed benchmark score measures a narrow slice of capability, and reaching it more cheaply does not automatically translate into better outcomes for users. The gap between laboratory metrics and practical reliability is a recurring theme in AI deployment, where error rates and edge cases often determine whether a tool is genuinely useful.
The combination of falling costs and rising capability also raises questions about competitive dynamics. If algorithmic progress continues at roughly three times per year, the advantage of scale may erode over time, allowing new entrants to challenge established players. But if reasoning models and other compute-heavy approaches dominate the frontier, the most capable systems may remain expensive even as the cost of yesterday's performance collapses.
For now, the data points to a technology improving on an unusually steep curve. The price of a fixed level of AI performance is dropping about 13 times annually, and even the underlying algorithmic progress is running at roughly three times per year. How those gains translate into everyday products, services, and prices will depend on how quickly developers pass savings on to users and how much additional computation the most advanced models demand.
4
