Every lab prices output tokens at 5–6× input. Why?

Put every frontier model’s input price on one axis and output price on the other, and the scatter collapses onto a line. That’s not a boring chart — it’s the finding. Across the models we track: Anthropic prices output at exactly 5.0× input on every current model ($10/$50 Fable, $5/$25 Opus, $1/$5 Haiku). OpenAI sits at 6.0× across its lineup, from the $0.20 nano tier to the $30/$180 pro tier. Google spans 5–8.3×, with the flash-lite tier carrying the market’s highest output premium.

Why the tight band? Output tokens are generated one at a time — sequential decode dominated by memory bandwidth — while input tokens are processed in parallel prefill. The compute asymmetry between the two phases lands, for current architectures and batching regimes, near that 5–6× band, and pricing has converged with it. Deviations are strategy: a high multiple on a cheap model taxes chatty workloads; a low multiple courts long-generation use.

Two practical consequences. First, comparing models on input price alone is a mistake the pricing structure invites — a “cheap” input rate with an 8× output multiple can out-cost a pricier peer on generation-heavy work; our tables carry a blended 3:1 column for exactly this. Second, when a lab moves off its ratio, that’s news — it signals a shift in serving economics. Our events feed catches every published change.

The live scatter — log-log, with the 5× and 6× rays drawn in — is at /models, and every price is verified against the lab’s official page with the verification date shown. First upcoming event we’re positioned to record: Claude Sonnet 5’s intro pricing is scheduled to end September 1, 2026 — a rare published price increase.