Right after a new AI skill tops the charts, its price crashes fast — then how long until the bargain era really kicks in?
What time horizon does the two-year performance price slowdown suggest?
This explores what Epoch AI's finding (prices for a new AI capability fall very fast at first, then slow down within about two years) tells us about how long a frontier capability stays expensive and exclusive before it becomes a cheap commodity.
This explores what the two-year slowdown in AI performance prices says about timing: how long a new capability stays scarce and pricey, and when it settles into ordinary, predictable cost declines. Epoch AI's starting finding is that when a model first reaches state of the art on a benchmark, the price of that level of performance falls about 66% per quarter. Within two years the decline slows to about 32% per quarter Do AI performance prices fall faster when capabilities debut?. Their explanation is economic, not technical. A new capability starts out with a temporary monopoly price, and competitors chip away at that premium as they catch up.
The percentages hide how big this is, so it helps to compound them. A 66% quarterly drop means the price is multiplied by about 0.34 every three months. That works out to roughly 75 times cheaper in the first year. A 32% quarterly drop compounds to about 5 times cheaper per year. That second figure is close to the 5–10x annual decline Epoch's Denain projects for reaching a fixed capability level through distillation, algorithmic gains and better hardware Does inference cost remain a persistent economic burden?. It also sits under the roughly 13x-per-year average across the whole period since 2023 How fast is AI capability becoming cheaper to produce?. So the corpus suggests two clocks. The first is a short premium window, measured in quarters, when early access to a capability carries a real price markup. The second is a long tail, measured in years, when cost falls at a steady engineering-driven rate. The corpus doesn't give the exact point where one clock hands off to the other. It only says the shift happens "within two years."
This has practical consequences. If you're deciding whether to pay frontier prices now or wait, most of the price drop happens in the first few quarters, so waiting even six months can change the economics completely. It's also why some researchers argue benchmarks should report token counts and not just dollar costs. A cost figure frozen at launch goes stale in exactly the window when prices move fastest Should benchmarks report token counts instead of frozen costs?.
There is a catch. A given capability gets cheaper, but the frontier itself doesn't. Benchmark performance per dollar improves 5–10x a year, while the absolute cost of running the newest frontier models rises 3–18x a year Is benchmark progress outpacing the cost to achieve it?. The two-year horizon applies to a fixed level of performance. If you always want the latest capability, you stay inside the expensive premium window and pay its price again and again.
An unexpected parallel: the "monopoly premium, then competition" pattern resembles Doctorow's platform lifecycle, but runs in the opposite direction. On platforms, value shifts away from users over time Do platforms inevitably decline through value extraction cycles?. In AI pricing, value shifts toward users as competitors catch up. A similar timing gap may affect how fast AI's economic payoff shows up. Brynjolfsson and colleagues argue that productivity gains look weak early because the supporting investment around a technology goes unmeasured Why does AI productivity appear weak when investment is high?. Put together, the period when a capability is cheap enough to adopt widely may arrive well before the period when its benefits show up in the numbers.
Sources 7 notes
Epoch AI found that across five benchmarks, performance costs drop 66% per quarter when newly state-of-the-art, then slow to 32% per quarter within two years. They attribute this to temporary monopoly pricing on novel capabilities that competition gradually erodes.
Denain argues the inference cost burden is not persistent because the dollar cost to reach a given capability level drops rapidly each year through distillation, algorithmic efficiency gains, and hardware advances. He documents a 3x cost reduction over eight months on a math benchmark and projects 5–10x annual reductions, suggesting today's expensive capabilities become affordable within years.
Epoch AI's five-benchmark analysis found AI capability prices dropping 13× per year since 2023—four times faster than DNA sequencing and 54 times faster than historical electricity cost declines. OpenAI's o3 achieved a benchmark result for $0.30 that GPT-5.6 Luna later matched for $0.0004.
Inference prices fall over time, making frozen dollar-cost figures misleading for procurement decisions. Reporting input/output token counts alongside costs lets downstream developers recalculate expenses as prices change, keeping benchmarks useful.
Regression analysis of frontier model pricing shows benchmark performance per dollar improves 5–10× annually, yet the absolute cost to run these models rises 3–18× annually. Roughly half of measured progress on GPQA-Diamond tracks increased inference spending rather than capability gains independent of compute.
Show all 7 sources
Doctorow identifies a three-phase lifecycle where platforms initially benefit users, then exploit business customers, then extract shareholder value. Amazon Marketplace, Facebook, and Twitter exemplify the pattern, though the research provides illustrative rather than sampled evidence.
Brynjolfsson, Rock, and Syverson show that GPT adoption generates large unmeasured complementary investments (training, processes, software) that initially undercount true productivity growth, then overcount it once those assets produce measurable output.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- The Price of Progress: Price Performance and the Future of AI
- The plunging price of thought
- How persistent is the inference cost burden?
- The rising costs of training frontier AI models
- Frontier AI capabilities can be run at home within a year or less
- Is a compute crunch coming?
- The Rise of Industrial AI in America: Microfoundations of the Productivity J-curve(s)
- Rise of agentic AI: How trust is the key to human-AI collaboration