The Price of Intelligence Is Falling Rapidly (marginalrevolution.com)

A recent report by Epoch AI highlights a dramatic decline in the cost of achieving specific levels of AI performance, dropping an astonishing average of 47% per quarter over the past three years. This trend represents a 13-fold reduction in costs annually, making it the fastest decline observed in any transformative technology to date. For instance, OpenAI's o3 model, which charged around $0.30 per question for 75% accuracy on the GPQA Diamond benchmark in January 2025, saw the emergence of GPT-5.6 Luna, which achieved similar performance for just $0.0004 in mid-2026—a staggering 725-fold decrease within an 18-month timeframe. This cost reduction is significant for the AI and machine learning community, as it means that not only are models becoming smarter, but they are also becoming considerably cheaper to operate, enabling broader accessibility and application. As performance levels improve with declining inference costs, the competitive landscape shifts: while new frontier models may still outperform their predecessors, the operational efficiency makes them less intimidating to legacy systems. This evolving dynamic suggests that innovation in AI/ML is not only accelerating in capability but also becoming more economically feasible, paving the way for wider adoption and experimentation in various sectors.