A recent analysis finds that the cost of achieving a fixed AI performance level — termed the "price of thought" by the authors — has fallen by roughly 13× per year over the last five years. The study argues this pace of decline outstrips comparable historical technologies: about four times faster than DNA sequencing, six times faster than declines in compute costs, 18× faster than lithium‑battery cost declines, and roughly 54× faster than electricity price reductions over the century up to 1973.
How the researchers measured it
The authors calculated the cheapest way to reach a given score on five AI performance benchmarks that cover mathematics, hard sciences, and game‑skill tasks (for example, chess) at multiple points over the past roughly three years. Their aim was to track how the minimum cost to attain the same performance level changed over time.
As an example, on January 31, 2025 OpenAI released a new iteration in its "reasoning" model series called o3. The authors estimate that, on average, achieving a 75% score on the GPQA Diamond multiple‑choice exam (which covers PhD‑level physics, chemistry and biology) with o3 cost about $0.30 per question. About 18 months later, OpenAI released GPT‑5.6 Luna, which achieved the same score on the same test at an estimated cost of $0.0004 per question. That represents a 725‑fold drop in the price of thought in under 18 months — analogously, the sticker price on a $50,000 car falling to $69 in 1.5 years.
The authors report similarly steep price declines on mathematics and chess benchmarks, and coarser data suggest the trend goes back to the start of commercial LLM inference when OpenAI fully released GPT‑3 in November 2021.
The measured rates of decline
Across the data, the cost to reach a given performance level fell about 47% per quarter on average, equivalent to roughly a 13× decline per year. Different task groups showed different speeds:
- Game‑based puzzles declined more slowly, about 39–43% per quarter (roughly 7–10× per year).
- Math problems showed faster progress, about 50–52% per quarter (roughly 16–19× per year).
The pace of decline tends to slow over time. When a performance level is first achieved, costs fall quickly — about 66% per quarter (≈75× per year). Two years later the drop rate is roughly half as fast, about 32% per quarter (≈4.7× per year). One possible explanation the authors offer is that firms can briefly charge a premium for newly achieved performance levels before competition and subsequent technical improvements drive prices down.
Caveats and where to look for the data
The study notes that benchmarks are an imperfect proxy for AI's real market costs or for direct comparisons to other historical technologies. The authors believe their numbers are reasonably representative of reality but caution they should not be treated as precise point estimates.
The full dataset and code are available on GitHub, and an overlay page accompanying the report provides numerous plots and tables for further exploration.
Conclusion
Over the past five years the estimated cost of AI "thinking" has collapsed at a rate unmatched by their reference technologies. Whether this rapid decline continues and what its broader implications for industry and the economy will be remain open questions that only future developments can answer.
