Epoch AI research measures AI inference cost declining approximately 13x annually after accounting for hardware efficiency gains, reaching unprecedented price-performance improvements.
The price of reaching a fixed performance level on select AI benchmarks has dropped sharply since 2023.
Epoch AI, a research organization that tracks AI trends, reports that costs are falling by approximately 47 percent per quarter on average—roughly 13x per year. No other transformative technology has declined that rapidly, the organization argues. This figure reflects market prices for a fixed benchmark score, however, not pure algorithmic or architectural progress, let alone real-world productivity costs.
MIT researchers analyzing comparable data observe costs dropping 5x to 10x annually. When they exclude cheaper hardware and competitive pricing pressure, the actual gain in algorithmic efficiency comes to roughly 3x per year. Peak performance per task can actually become more expensive, since newer reasoning models require substantially more compute per question.
The two studies answer different questions: matching last year's top-tier capability costs far less, but running today's best model often costs significantly more per query.
Epoch uses OpenAI's o3 as an example. In early 2025, o3 scored 75 percent on GPQA Diamond, a PhD-level science test, at an estimated 30 cents per question. Eighteen months later, a GPT-5.6 family model achieved the same score for 0.04 cents. Epoch describes this as 1/725 of the original price. If automobile prices fell at this rate, a 50,000-euro vehicle would cost less than 70 euros. OpenAI launched the cheaper GPT-6 Sol and Luna models only days ago, suggesting the gap has widened further.
Epoch bases its analysis on five benchmarks spanning mathematics, science, and logic puzzles. Acknowledging this narrow sample, the organization describes its findings as "reasonable but rough measurements based on the best available data."
Hans Gundlach and his MIT colleagues draw on pricing data from the comparison platform Artificial Analysis, covering April 2024 through November 2025, and evaluate substantially more models per test than prior research. Their figures run lower than Epoch's partly because newer reasoning models deploy additional test-time compute to tackle harder problems, raising the cost per correct answer even as per-token prices decline. Since tokens are the text units providers charge for, per-token comparisons alone capture an incomplete picture.
When MIT breaks down the drivers of falling prices, cheaper hardware accounts for some of it, and competition for the rest. By examining open models separately, the researchers isolate the pure algorithmic efficiency gain: approximately 3x per year. Epoch's 13x figure is higher because it does not exclude hardware and competition effects.
MIT also found that some performance improvements come simply from allocating more compute. A new model that outperforms its predecessor on GPQA Diamond appears to represent progress, but the authors estimate that a substantial share of the improvement stems from greater computing resources per question. It scores higher and costs more to run. Coding and math benchmarks show a smaller version of the same dynamic.
Not all benchmark progress reflects genuine efficiency gains. New models combine better training, better data, better architecture, and more test-time compute. A single score blends all of these factors. There is also the "benchmaxxing" problem: AI companies might optimize for well-known tests, inflating scores without real-world benefit. Epoch attempts to guard against this by including one test, "Mystery Game Puzzles," based on a secret game. Costs drop slowest on that test, consistent with benchmaxxing theory, though the pattern could equally reflect task format or data noise.
Price alone will not determine which model to choose. Broad cost averages offer little guidance when selecting a model for a specific task. Platforms like Artificial Analysis rank models across quality, price, latency, context window, and output speed, and the cheapest option rarely excels across all dimensions. A low-cost model with high latency is impractical for real-time chatbots. A powerful reasoning model may be too slow for automated workflows. A more expensive frontier model could still reduce total costs if it succeeds more often and eliminates the need for retries. None of this appears in a simple price-per-token calculation. We explored this token economics question in detail in Frontier Radar #3.