Your AI subscription is a meter you can't see
12 Oct 2026 · News
The price on an AI plan stays the same every month. What it buys does not, and the people paying the most are the first to notice.
The flat monthly AI subscription is now a label on a metered plan. Within two quarters, we expect every major lab to publish a real credit schedule for its top consumer tier, leaving the roughly $20 plan as the last one sold as a plain flat price.
Start with the clearest case. On 29 September OpenAI cut its $200 Pro plan from 20 times the Plus allowance to 10 times, and added a $500 tier at 25 times, according to The Next Web, which cites OpenAI's help page. Existing subscribers keep the old limits until 29 October. Same price, half the plan.
Anthropic is more candid about the machinery. Its pricing page, read today, says: "Every plan has usage limits that reset on a rolling five-hour session window, and paid plans add weekly limits on top." Run out and you can "turn on usage credits to keep working at standard API rates." The top plan is sold as 5 or 20 times "more usage than Pro." That is a multiple of a number nobody publishes. A meter you can't see.
Why now? SemiAnalysis, an industry research firm, estimated on 5 October that subscriptions are about 10% of Anthropic's revenue but "can take up over 40% of inference compute." That is their model, not a company disclosure. They also bought three identical subscriptions and found one had about 20% lower limits. The provider called it an "extremely tiny" A/B test. SemiAnalysis's conclusion: "providers can silently change subscription limits at any time."
The reason a flat price stops working is that users are nothing alike. In a16z's latest consumer AI report, the top 1% of payers accounted for 19.5% of observed consumer AI spend, more than the bottom half combined (16.6%). That top 1% averaged $903 a month. The median payer spent $25. (All of this sits inside a small group: only 4.5% of eligible consumers in YipitData's panel paid for ChatGPT, Gemini, or Claude in August, up from 2.1% a year earlier.) Cursor, the AI coding app, found the same shape in its own usage: over four weeks, its top 10% of users consumed "nearly two-thirds of all tokens."
Put those together and the flat plan looks like an average that fits almost nobody. The median user is cheap to serve and fits comfortably under $20. The heavy few burn the compute, and they are exactly who the $100 to $500 tiers are for.
Meanwhile, the bottom of the curve keeps getting cheaper. Anthropic's Haiku 5.5, its small model, launched 7 October "priced 90% lower than Claude Haiku 4.5" for requests up to 100,000 tokens. The same post gives Max subscribers $100 or $200 a month in API credits. Read that as a tell: part of the top tier is already paid out in metered units. And a16z counts 64% of AI-native products in its top web rankings already selling usage charges or extra credits.
So the direction is settled. Cheap, light use gets cheaper and stays flat. Heavy use gets a meter. The open question is only whether the meter gets a visible dial.
Our call: by the end of April 2027, OpenAI, Anthropic, and Google will each publish, for their most expensive consumer plan, an actual unit (credits per model, tokens, or dollars at API rates) instead of "N times more than the plan below." What would prove us wrong: a major lab launching a genuinely unlimited top tier that holds for a full quarter.
What this desk checked: we read Anthropic's pricing and Haiku 5.5 pages today, the a16z report, Cursor's post, and the free portion of SemiAnalysis (the rest is paywalled). OpenAI's help pages blocked our fetch, so the Pro numbers rest on reporting that cites them. We did not check Google's plans or test anyone's limits ourselves.
The price tag is a flat number. The thing behind it is a meter, and the labs are already reading it. Soon they will have to show it to you too.