Cheaper tokens won't shrink your bill. They'll change what you build.

On Monday, September 22, 2026, Anthropic dropped the price of Claude Opus 5.5 — output tokens fell to around $20 per million, down from $25, roughly a fifth cheaper, on a model it says outruns its own larger one. Within hours OpenAI answered with two GPT-6 models, Sol and Luna, at something close to half their predecessors’ prices; Luna’s output reportedly went from about $1.20 per million to $0.50. Simon Willison called it what it was: a price war. If you build on these APIs, your first thought was probably a happy one — my bill is about to go down.

It almost certainly isn’t.

There’s a 160-year-old idea that saw this coming. In 1865 the economist William Stanley Jevons noticed something strange about coal: as steam engines got more efficient and used less coal per unit of work, England burned more coal, not less. Cheaper steam made whole new uses worth it — factories, railways, ships that never would have paid at the old price. Efficiency didn’t shrink demand. It unlocked it. Today we call it the rebound effect, or Jevons paradox, and it is about to run straight through your cost model.

Swap coal for tokens. The call you would never make at $20 per million — re-rank every search result through a model, generate three drafts and keep the best, run a verifier pass over every write, let an agent retry until it’s confident — is a rounding error at fifty cents per million. So you make it. Constantly. Usage expands to fill the lower price. The unit cost collapses and the invoice stays flat, or climbs, because you are now doing things that were uneconomical last week.

That reframes what a price cut actually is. Don’t book it as savings to bank. Book it as a capability budget to spend. The question a cheaper token asks you is not “how much will I save?” but “what would I build if this call were free?” — because it is getting close, and your competitor is answering the same question tonight.

Two things follow. First, the moat is never the cheap token. Everyone gets the same price on the same day — that’s what a price war is. Your advantage is what new behavior you wire the cheap token into first, and whether that behavior compounds into something a price sheet can’t copy. Second, a falling floor hides waste. When calling the biggest model for everything cost real money, the bill disciplined you; it was a signal. When it’s nearly free, sloppy defaults stop stinging — until scale turns a rounding error back into a line item. Cheap is not free, and “nearly free × enormous volume” is exactly how a surprise invoice is born.

The reflex to unlearn: a price drop is not the market handing you margin. It’s the market moving the floor so a new class of product becomes buildable — and quietly obligating you to build it before someone else does.

The idea, to look up: Jevons paradox and the rebound effect — William Stanley Jevons, The Coal Question (1865); modern energy-economics work on direct vs. indirect rebound. The rebound is often partial and occasionally exceeds 100% (“backfire”), and it’s always context-dependent — treat the direction as robust and the magnitude as an open question.

The news, to check: Anthropic’s Claude Opus 5.5 price cut and OpenAI’s GPT-6 Sol / Luna launch, both Sept 22 2026 — TechCrunch (Russell Brandom), Simon Willison’s “Claude Opus 5.5, GPT-6 Sol and Luna, and a new price war,” and SiliconANGLE. Prices as reported and moving fast; verify current API rates before quoting any figure.