A report by Epoch AI indicates that the cost of artificial intelligence capabilities has dropped by hundreds of thousands of times over the past five years. This decline is occurring at a rate of nearly 50% per quarter, or 13 times cheaper annually, significantly outpacing historical cost reductions for lithium batteries and DNA sequencing.
The analysis focused on models achieving at least 81.25% on the GPQA Diamond benchmark. For instance, the cost to answer a question on this test fell from $0.30 with OpenAI's GPT o3 in January 2025 to $0.0004 with the GPT-5.6 Luna model released eighteen months later.
Despite the rapid decline, the pace of cost reduction is slowing. Initial data showed a 66% drop per quarter, which decreased to 32% per quarter two years later. Factors such as rising energy prices and compute hardware costs in 2026 may be contributing to this trend.
The report warns of 'benchmaxxing,' where developers optimize models specifically for test performance. While Epoch AI used randomized elements to mitigate this, the study acknowledges that some benchmark optimization remains likely, suggesting the results should be viewed with a degree of skepticism.
For companies like OpenAI and Anthropic, these falling costs present a business challenge. As intelligence becomes cheaper and more accessible, maintaining customer loyalty and consistent revenue becomes difficult, as users may frequently switch to newer, more affordable alternatives.
