AI 性能成本暴跌，但别急着下结论
论证要点
- Epoch AI 测算，自 2023 年以来达到固定基准分数的成本平均每季度下降约 47%，约合每年 13 倍，速度超过以往任何变革性技术。
- MIT 研究者用 Artificial Analysis 的定价数据得出更低数字：年降 5 到 10 倍；剔除硬件降价和竞争压力后，纯算法效率提升约为每年 3 倍。
- 两项研究口径不同：复现去年顶级能力变得极便宜，但运行当前最强模型往往每次查询更贵，因为新推理模型会消耗更多测试时算力。
- Epoch 以 o3 为例：2025 年初在 GPQA Diamond 上得 75 分约需 30 美分一题，18 个月后同分只需 0.04 美分，约为原价的 1/725。
- MIT 发现部分基准进步只是花了更多算力：新模型分数更高但单题成本也更高，编码和数学基准也呈现类似但较小的模式。
- 存在“benchmaxxing”风险：厂商可能针对知名测试优化，Epoch 用保密游戏 Mystery Game Puzzles 防范，该测试成本下降最慢，但作者也承认可能是任务格式或数据噪声所致。
对实践的启示
对做 AI 应用的人来说，复现旧能力确实越来越便宜，但选型不能只看每 token 价格。延迟、上下文、输出速度和重试成本都会影响总账，最便宜的模型很少在所有维度都赢。
作者承认的反方与局限
Epoch 承认其结论基于五个基准的狭窄样本，只是“合理但粗略”的测量；成本下降最慢的保密测试既可能印证 benchmaxxing，也可能只是任务格式或数据噪声。
原文摘引
No other transformative technology declined that fast, the group says.
Matching last year's top-of-the-line capability? Dramatically cheaper. Running the current best model? Often significantly more per query.
Not all benchmark progress is efficiency progress.
▸ 展开原文全文（30 段，抓取自 The Decoder）
AI performance costs are falling faster than those of any previous technology
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, says costs are falling by about 47 percent per quarter on average, or about 13x per year. No other transformative technology declined that fast, the group says. The number reflects market prices for a fixed benchmark score, though, not pure algorithmic or architectural progress or even real-life productivity costs, which is a whole other story.
MIT researchers looking at comparable data see costs dropping 5x to 10x annually. Once they strip out cheaper hardware and competitive pricing pressure, they put the actual gain in algorithmic efficiency at about 3x per year. Peak performance per run can actually get pricier, because newer reasoning models burn through a lot more compute per task.
Both studies are asking different things, though. Matching last year's top-of-the-line capability? Dramatically cheaper. Running the current best model? Often significantly more per query.
Matching o3's accuracy now costs a fraction of the price
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 hit the same score for four hundredths of a cent. Epoch says that's 1/725 of the original price. If cars dropped that fast, a 50,000-euro vehicle would cost less than 70 euros. OpenAI launched the even cheaper GPT-6 Sol and Luna models just days ago, so the gap has likely widened further.
Epoch bases its analysis on five benchmarks spanning math, science, and logic puzzles. Because that's a narrow sample, the organization calls its findings "reasonable but rough measurements based on the best available data."
Algorithmic gains only explain part of the drop
Hans Gundlach and his MIT colleagues use pricing data from the comparison platform Artificial Analysis, covering April 2024 through November 2025, and evaluate far more models per test than previous research. Their numbers are lower than Epoch's partly because newer reasoning models can throw extra test-time compute at hard problems, driving up the cost per correct answer even while per-token prices keep falling. Tokens are the text units providers bill for, and comparing them alone misses the full picture.
When MIT breaks down what's pushing prices lower, cheaper hardware accounts for some of it, competition for more. The researchers control for competition by looking at open models separately. What's left is the pure algorithmic efficiency gain, which comes out to about 3x per year. Epoch's 13x figure is higher because it doesn't strip out hardware and competition effects.
Better benchmark scores don't always mean better efficiency
MIT also found that some performance gains simply come from spending more compute. A new model that beats its predecessor on GPQA Diamond looks like progress from the outside, but the authors estimate a big chunk of the improvement comes from using more processing power per question. It scores higher and costs more to run. Coding and math benchmarks show a smaller version of the same pattern.
Not all benchmark progress is efficiency progress. New models roll together better training, better data, better architecture, and more test-time compute. A single score blends all of that.
There's also the "benchmaxxing" problem: AI companies could optimize for well-known tests, inflating scores without real-world payoff. Epoch tries to guard against this by including one test, "Mystery Game Puzzles," based on a game that's been kept secret. Costs drop slowest on that test, which fits the benchmaxxing theory but could just as easily reflect the task format or noise in the data.
Price alone won't tell you which model to pick
Broad cost averages don't help much when you're choosing a model for a specific job. Platforms like Artificial Analysis rank models across quality, price, latency, context window, and output speed, and the cheapest option rarely wins on every dimension.
A low-cost model with high latency is useless for a real-time chatbot. A powerful reasoning model might be too slow for automated workflows. A pricier frontier model could still save money if it gets things right more often and cuts down on retries. None of that shows up in a simple price-per-token comparison. We dug into this token economics question in Frontier Radar #3.
AI News Without the Hype – Curated by Humans
Subscribe now
Read on for the full picture.
Subscribe for hype-free coverage.
- Full access to every article on THE DECODER
- No ads
- Join the comments and community discussions
- A weekly AI news recap via mail
- 6x/year: "AI Radar" — deep dives on the AI topics that matter most
- Daily AI news, always up to date
- Our full ten-year archive
- Covered by a team with 10+ years in AI