The Cost of Reaching a Fixed AI Performance Level Is Dropping Fast

The cost of achieving a fixed level of AI performance is falling roughly 13-fold per year, according to Epoch AI — a rate of decline faster than that of any previous technology.

The finding comes from Epoch AI, a research organization that tracks progress in machine learning. It measures how the price of reaching a set benchmark performance level has changed over time. The answer: dramatically. A model capable of hitting a given score on a benchmark today costs about one-thirteenth of what an equivalent model cost a year earlier.

A separate analysis from MIT adds nuance to that headline figure. After stripping out the contribution of hardware improvements and market competition, MIT researchers estimate the underlying rate of algorithmic progress at about 3x per year. In other words, better algorithms alone triple the cost-efficiency of AI systems annually. The rest of the 13x decline comes from cheaper, faster chips and competitive pricing pressure among providers.

Why the numbers matter

The stakes here are economic and strategic. If the cost of a given capability drops by an order of magnitude every year, capabilities that were prohibitively expensive for most organizations become commodities within a few budget cycles. That compression shapes who can build with AI, at what scale, and how quickly AI-heavy products reach mass-market prices.

The rate also outpaces historical precedents. Previous general-purpose technologies — electrification, computing hardware, solar photovoltaics — improved on cost curves measured over decades, not single-year multiples of this size.

A caveat on today's frontier models

The falling cost curve does not mean the best current models are getting cheaper to run. Epoch AI's finding applies to reaching a fixed performance level, not to the price of frontier systems. Reasoning models illustrate the distinction: they can cost more than their predecessors because they consume far more compute per task, spending extended inference time working through problems step by step.

The practical implication for buyers is straightforward. Price is only one variable. When selecting a model for real-world deployment, quality, speed, and error rate matter just as much as cost. A cheaper model that fails more often or responds too slowly can be the more expensive choice in production.

The dynamic creates a two-track market: rapidly commoditizing mid-tier capabilities, and increasingly expensive frontier reasoning systems that push past current limits. For enterprises and developers, the question is less whether AI costs will fall — Epoch AI's data suggests they will, and quickly — but which tier of performance a given application actually needs.