The cost of using AI models is undergoing a rapid drop; a trend that could seriously change the advanced software market.
According to ITNA quoting TechSpot, a new report from Epoch AI shows the cost of achieving a given level of AI capability has fallen by about 50 percent each quarter; meaning the cost of reaching the same performance level is roughly 13 times cheaper year over year.
Instead of examining average prices of AI products, the report measures the cost for models to reach a specified level of performance. One of the main metrics is the GPQA Diamond test, which includes graduate-level science questions.
The difference between new and old models clearly shows the pace of this cost decline. According to the report, the GPT o3 model released in January 2025 cost about $0.30 per question to reach a 75 percent score on this test. About a year and a half later, GPT-5.6 Luna could deliver the same level of performance at a cost of about $0.0004 per question.
Epoch AI compared this trend to other technologies such as lithium-ion batteries, computing, and DNA sequencing. Based on this comparison, the cost of AI capabilities has fallen about four times faster than DNA sequencing and 18 times faster than lithium batteries. Of course this comparison has limitations; because time spans and calculation methods are not the same across technologies.
Also, cost reductions are not the same across all AI tasks. Reviewing metrics such as mathematics and chess problems has shown that some tasks experience gradual cost declines, while others go through long periods of stability and then sudden drops.
The pace of cost decline also appears to be slowing. The average cost reduction across five reviewed metrics was initially about 66 percent per quarter, but two years later it fell to 32 percent. Energy costs, datacenter capacity, and AI hardware are factors that can make further cost reductions more difficult.
The report also warns about the phenomenon of "benchmark optimization" or benchmaxxing; that is, optimizing a model to score highly on tests without necessarily improving its real-world performance to the same extent. Despite these limitations, cheaper AI could give companies more choices and put greater pressure on providers of advanced models.