The Bleeding Edge

// Explainer

The price of intelligence is collapsing

What a million tokens of frontier-model output cost at launch, 2020 to today — and why the collapse, not the capability, is the real story.

Launch-era list prices, $ per million output tokens. Toggle tiers · switch scales · hover points for detail.

So what?

In September 2020, a million tokens of the best available model cost $60. Frontier launch prices have stayed in the same band ever since — GPT-4 at $60, Claude 3 Opus at $75, o1 at $60. The frontier itself is not getting cheaper; it prices as a premium product on the day it ships.

The collapse happens one tier down. Capability that was frontier-priced becomes near-free within roughly eighteen months: GPT-4-class output went from $60 to $15 (GPT-4o) to $0.60 (GPT-4o mini) in under two years, and DeepSeek-V3 shipped frontier-class output at $1.10. Switch the chart to linear scale to feel how violent that drop is — on the log scale it reads as a slope, on linear it reads as a cliff.

The applied takeaway: if a use case almost works economically today, it will work in about a year without you doing anything. Price your product on where the curve is going, not where it is.

// Sources

Launch-era list prices per million output tokens; the linked pages may show newer prices today. 2026 models appear once their launch prices are added with sources.