The $200 AI seat is a subsidy: ceiling users pull 40x to 70x what they pay

Forty times and seventy times: those are the ratios once the API-equivalent figures are divided by the $200 price ^[1][2]. Held for a year, the ceiling user on the Anthropic plan draws about $96,000 of list-rate inference for $2,400 in fees, and on the OpenAI plan about $168,000 ^[3][4].

Most subscribers will never get near that ^[7]. The researchers got there by running one long workload continuously against the weekly caps rather than holding conversations ^[3], and agentic coding is precisely the workload that behaves that way, consuming far more tokens than chat or a one-shot request ^[14]. A team leaving agents running overnight is not an unusual user. It is the test subject.

Prompt caching is the one mitigation with a published number attached: an independent study cited in the research cut effective API cost by 88.6%, to roughly $0.57 per million tokens ^[10]. Apply that discount to the ceiling and the Anthropic case still represents about $912 a month of wholesale inference on a $200 plan, the OpenAI case about $1,596 ^[5][6]. Measured against its own best fix, the subsidy is still 4.6x to 8x ^[7]. Caching narrows the gap. It does not close it.

Price is not the only dial available. The $200 tiers are sold as 20x the entry plan's usage, with $100 tiers at 5x ^[6], and that is a quota, not a dollar amount. It can be redefined without anyone announcing a price rise, and it does the same work on margin.

There is also evidence buyers will not volunteer to cover the difference. Ramp's August AI Index put Anthropic ahead of OpenAI in US business adoption, 43.5% of companies to 39.7% ^[11], but Ramp lead economist Ara Kharazian has noted that Anthropic's most expensive model took just 6% of the tokens customers bought in its first month ^[12]. Adoption is broad; appetite for the top of the menu is not. That is the demand curve the labs have to reprice against, and it argues for quieter limits ahead of visible list prices.

SemiAnalysis frames its finding as a margin question, whether flat-rate plans hold without tighter caps or higher prices ^[15]. The buyer's version is duller and more useful: the durable number is not $200 a seat, it is the multiple sitting behind it, and a headcount plan built on today's multiple contains an input nobody has priced. The 30x spread in token use across identical tasks ^[9] means that input cannot be estimated from your own usage history either.