💸 You’re paying for tokens. Now what?
Exponential View Azeem Azhar
AI companies are ditching flat subscriptions for pay-per-token pricing, especially on coding tools. Heavy users who were quietly costing labs a fortune are about to feel it in their wallets.
Based on reporting by Exponential View, Azeem Azhar — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Think about your gym membership. You pay the same flat fee whether you show up twice a month or ten times, and the gym is fine with that because serving you barely costs them anything extra. AI companies used to run on the same logic. ChatGPT and Claude subscriptions worked like all-you-can-eat buffets, and most people used them the way most people use gym memberships: occasionally, guiltily, in January.
But a chatbot query and an AI agent chewing through code are not the same kind of visitor. OpenAI's finance chief Sarah Friar has said Pro subscribers hammer ChatGPT roughly 11 times more than free users. That gap is already big for a person typing questions by hand. It gets absurd once you hand the keyboard to an agent. A human fiddling with a chatbot might struggle to burn through 100,000 tokens in a day. An autonomous coding agent left running will eat 100 million tokens before lunch. One user, apparently, racked up 130 billion tokens in a single month.
That asymmetry is why the bundled-pricing era is ending, at least for coding tools. Uber just capped its 5,000 developers at $1,500 a month, or $18,000 a year, per agentic coding tool, with room to request more if they need it. Run the math and the worst case is a $90 million annual bill against Uber's $9.8 billion in 2025 free cash flow. Less than one percent. This isn't really about companies being unable to afford AI. It's about whether they can tell if the spending is actually producing value, and metered pricing forces that question into the open in a way flat bundles never did.
There's a useful historical parallel here, and it's internet advertising. Back in 1994, when AT&T ran that first banner ad on HotWired, ads were sold in bulk, cost-per-thousand-impressions, click or no click, value or no value. Eventually the industry shifted to pay-per-outcome pricing, and rather than shrinking the market, that shift blew it wide open. Advertisers stopped paying for maybe-attention and started paying for results, and spending followed. AI pricing looks headed down the same road: less generous on the surface, but potentially a bigger, healthier market underneath, once buyers can actually see what they're getting for their money.
My take — AI-written commentary, not fact-checked reporting
I've said for a while that flat AI subscriptions were a temporary illusion propped up by venture money, not a sustainable business model, and this is the correction arriving on schedule. Metered pricing isn't the labs getting greedy, it's the market finally pricing agentic AI according to what it actually costs to run, and that's healthy even if it stings power users. The advertising precedent is the right one to watch: pay-per-outcome models tend to grow markets, not shrink them, once buyers can see the ROI clearly instead of guessing inside a bundle.
Read more about this at: Exponential View
