AI's Cost Plummets Dramatically, Outpacing All Prior Technologies

Artificial intelligence is undergoing a revolution not just in capability, but in affordability, with its cost reduction rate outpacing every major transformative technology in history. New analysis reveals that the expense associated with achieving a specific level of AI performance has been plummeting by roughly 47% each quarter since 2023. This translates to a staggering thirteen-fold decrease in cost annually, a pace that dwarfs historical benchmarks like DNA sequencing, general compute power, and even electricity.

The Unprecedented Pace of AI Cost Reduction

The "GPT" in OpenAI's models, standing for Generative Pre-trained Transformer, hints at the broader ambition of these technologies as general-purpose tools. Like the steam engine or the internet, AI is poised to reshape society. While the macroeconomic impact of AI, driving up demand for inputs like chips and power, is widely recognized, the equally significant, yet paradoxical, decline in the cost of AI's output is less appreciated. For instance, a reasoning model released in early 2025 could achieve a 75% score on a PhD-level science exam for about 30 cents per question. Less than 18 months later, a successor model achieved the same score for an astonishingly low $0.0004 per question—a 725-fold price reduction.

This dramatic cost decrease is not uniform across all AI tasks. Performance improvements on game-based puzzles, for example, have seen slower price drops of 39-43% per quarter. In contrast, AI's proficiency in mathematical problems has driven faster cost reductions, with prices falling 50-52% per quarter. The analysis, which spans three years and includes benchmarks in mathematics, hard sciences, and games, highlights that the most rapid cost declines often occur immediately after a new performance level is first achieved, when it represents the state-of-the-art.

Frontier Economics and Methodological Nuances

When a performance level is first attained, AI companies can command a premium. However, competition and ongoing technological advancements quickly drive these prices down. Averaging across five key AI capability benchmarks, the cost for newly achieved state-of-the-art performance plummets by 66% per quarter (75 times per year). Two years later, this rate slows to approximately 32% per quarter (4.7 times per year). This dynamic suggests an initial period of rapid price erosion followed by a more moderate, though still significant, decline.

It is crucial to acknowledge the caveats accompanying this analysis. The focus on the "frontier"—the cheapest model capable of a given performance level—assumes users are constantly seeking the most cost-effective solutions, which may not reflect real-world user behavior. Furthermore, the data, while comprehensive, is still relatively short-term (three years) and may not capture all model-benchmark combinations. The potential for "benchmarking"—training models specifically to excel on tests rather than general tasks—also introduces a degree of noise. Despite these limitations, the researchers believe the overall trends are representative of the rapidly evolving AI cost landscape.

Historical Context and Previous Research

This is not the first study to document the falling cost of AI. Previous analyses, such as those from Andreessen Horowitz in 2024 and Epoch AI in 2025, have indicated significant cost reductions in Large Language Models (LLMs) following the release of GPT-3. These earlier studies often focused on price per token. However, the advent of "reasoning models" has made it more critical to analyze the actual cost to achieve a specific performance level, rather than just token efficiency. More recent work, including an analysis on LessWrong and a detailed paper by Gundlach et al. in March 2026, has further refined these cost-performance metrics, finding annual declines ranging from 5x to 10x.

The methodology employed in the latest analysis is novel in its capture of each model's full expense-performance continuum. By mapping the "Pareto frontier"—the most cost-effective way to achieve each performance level—the study provides a more nuanced view than simply looking at peak performance. This involves predicting performance under various budget constraints by analyzing benchmark run transcripts, a technique adapted from the federal Center for AI Standards and Innovation (CAISI). This approach allows for a more accurate representation of the cost-performance trade-offs available to users.

Implications for the Future of AI

The implications of AI becoming dramatically cheaper are profound. It democratizes access to advanced computational capabilities, enabling smaller companies, researchers, and even individuals to leverage sophisticated AI tools. This could accelerate innovation across countless fields, from scientific discovery and drug development to personalized education and creative arts. As the cost of "thought" continues to fall, the economic barriers to deploying AI solutions will diminish, potentially leading to widespread adoption and transformative societal changes. The rapid commoditization of AI performance suggests a future where advanced AI is not a luxury but a ubiquitous utility, driving unprecedented productivity gains and new forms of economic activity.

While the exact figures may vary, the overarching trend is undeniable: AI is becoming astonishingly cheaper, faster than any technology before it. This trajectory promises to reshape industries and redefine what is possible, making the "price of thought" a critical metric for understanding the future of technology and the global economy. The ongoing research and development in this area will continue to push the boundaries of both AI capabilities and their economic accessibility.