From $0.30 to $0.0004 a Question: OpenAI's 725x Cost Collapse Shows AI Intelligence Is Entering Its Big Price-Cut Era

The AI industry is presenting a rather paradoxical picture. Chips, electricity, and data centers keep getting more expensive — even the electricians needed to build data centers are getting harder to hire — while at the same time, the cost of getting AI to complete a task of a given difficulty is falling at an unprecedented pace.

Epoch AI analyzed five categories of benchmarks over the past three years — covering math, science, and games, among others — and found that the cost required to reach the same level of model performance is dropping by about 47% per quarter on average, which works out to roughly 13 times cheaper every year.

That pace of price decline is 4 times faster than DNA sequencing, 6 times faster than computing costs, 18 times faster than lithium batteries, and 54 times faster than electricity — possibly faster than any general-purpose technology in human history.

A more tangible example comes from OpenAI. In early 2025, o3 reaching a particular level of performance on GPQA Diamond cost an average of $0.30 per question.

Less than 18 months later, GPT-5.6 Luna reaches the same level at just $0.0004 per question — a 725-fold drop in cost.

The closer to the frontier, the faster prices fall

The more remarkable pattern is that the closer a capability sits to the model frontier, the faster its price typically drops.

When a capability has just become state of the art, the associated cost falls by an average of 66% per quarter — about 75 times cheaper within a year. Only after about two years does the decline gradually slow. A capability billed as a flagship selling point today may move into small models within months, and then become an ordinary option in cloud services.

That also explains why large-model companies keep releasing new models, cutting prices, and then releasing newer models.

Training frontier models still requires ever-growing capital investment, but the window in which leading capability can command premium prices keeps shrinking. Each breakthrough briefly creates a price premium; competitors, open-source models, and engineering optimizations then quickly erase the gap.

Only two effective plays are left

By this year, competing on price-performance has become an industry consensus, and model launches increasingly resemble just two viable strategies.

Either justify a high price with a generational leap that punches through the ceiling, or push already-mature capabilities to a low enough price and redraw the market's "kill line" with extreme price-performance.

Standing in between is becoming harder and harder.

Without clearly leading capability and without a sufficient price advantage, a model — even with respectable leaderboard scores — can quickly lose its presence in today's dense release cycles.

Of course, benchmark results do not equal real-world performance, and users will not hunt for the cheapest model for every task. Epoch AI itself cautions that some models may have been trained toward the benchmarks, and three years of data is not enough to precisely characterize every trend.

Even so, the direction is hard to ignore: AI's "production side" keeps getting more expensive, while its "output side" keeps getting cheaper.

Over the past few years, the industry has defined competition through bigger parameter counts and higher leaderboard scores. But the more important question going forward may be who can commoditize rapidly depreciating model capabilities the fastest.

Models increasingly resemble electricity: what will truly be valuable is the workflows, data, user relationships, and delivery capabilities built around them. Intelligence will eventually be free; value still has to be created.