Nobody fell in love with AI. They fell in love with AI below cost.
The flat-rate AI era was a subsidy, and the bill is now landing in public. A look at the receipts from Cursor, GitHub Copilot, OpenAI, and Anthropic.
The story we keep telling is that AI went mainstream because it was good. That is half true. The other half is that for about three years, the average person was using frontier models at a price that had almost nothing to do with what the model cost to run. Writing, coding help, research, the daily “just ask the thing”: all of it felt close to free because someone else was paying the difference.
That was never going to hold. Inference costs real money, the companies eating those losses raised money on the promise of eventually not eating them, and the part of that cycle where they stop is not a prediction anymore. It is happening in public, on pricing pages, right now.
Here are the receipts.
The subsidy was real, and they admitted it
You do not have to infer the subsidy. The people running these companies have said it out loud.
In January 2025, Sam Altman posted that OpenAI was losing money on its $200-per-month ChatGPT Pro plan, saying people used it much more than expected and that he had personally picked the price thinking it would turn a profit. Read that again: the most expensive consumer tier, the one aimed at the heaviest users, did not cover its own compute. The cheaper tiers were never going to either. OpenAI’s single largest cost is the computing power to run the model, and that cost scales with use.
The macro numbers say the same thing louder. In the first half of 2025, OpenAI reportedly brought in about $4.3 billion in revenue while posting a net loss in the billions, with compute as the dominant line item (The Register, citing The Information). A business does not absorb losses at that scale forever. It either changes the price or changes who pays.
The roll-off is already on the pricing pages
This is the part that makes the “hangover is coming” framing outdated. It is not coming. Watch what the coding tools did, because developers are the canary: they are the heaviest users, on the most expensive agentic workflows, so they hit the real cost first.
Cursor. In June 2025, Cursor moved from a clean request-based plan to usage-based credits, where your monthly allowance depletes at the underlying model’s API rate. The $20 Pro plan now includes roughly $20 of frontier-model usage and then bills overages at cost. The rollout went so badly that the company issued a public apology on July 4, 2025 and refunded surprise charges. Independent write-ups noted the effective number of premium requests on the same $20 plan dropped from around 500 to around 225, and that for long-context agentic work the effective price could jump 20x or more. Same subscription, same headline number, very different bill.
GitHub Copilot. This is the strongest receipt, and it is three days old. GitHub started billing for premium requests on June 18, 2025, capping Pro at 300 a month with $0.04 overages. Then, on June 1, 2026, all Copilot plans moved to fully usage-based billing measured in tokens, input, output, and cached, at listed API rates. GitHub’s own words for why: to get to a sustainable, reliable business. Coverage of the change reported the fallback free model was removed and that power users on agentic sessions are projecting 10x to 50x higher bills. Two structural pricing changes on 4.7 million paid subscribers in under twelve months, both in the same direction.
Anthropic. On August 28, 2025, Anthropic added weekly rate limits to Claude Pro and Max, on top of the existing five-hour caps, explicitly to rein in people running Claude Code continuously. Hit the weekly cap and you buy more at standard API rates. The company said it affects under 5% of subscribers, which is the tell: the heaviest 5% were consuming an unsustainable share, so the price had to find them.
Three different companies, three different mechanics, one identical move: stop pretending compute is free, and route the real cost back to whoever is generating it.
Who actually pays when the price is real
Here is the uncomfortable half of the thesis. When the price reflects cost, most people do not pay it.
Across ChatGPT’s roughly 800 to 900 million weekly users, about 5% pay for a subscription, according to a senior OpenAI executive who spoke to the Financial Times. Menlo Ventures put the number for generative AI overall even lower: around 3% of the 1.8 billion people using these services pay anything at all. Almost 90% of ChatGPT’s users are outside the US and Canada, where they are worth far less to advertisers, which is why the cheaper regional tiers and the coming ad-supported plans exist.
So the “AI changes everything for everyone” line was always two things stacked together: a genuinely useful product, and a subsidy that let everyone treat it as free. Pull the subsidy and you get a split. A small group keeps paying because the value is obvious and the math works. Everyone else drifts down to a weaker free tier, an ad-supported tier, or back to doing the thing the old way.
What survives full price
The use cases that survive are the ones where the output has a clear dollar value, and usually where someone other than the end user is paying.
Coding survives, which is exactly why the coding tools were first to meter: a tool that saves an engineer real hours is easy to justify at API cost. Anything billed by the hour survives, because the AI is displacing labor that already had a price. Narrow vertical tools survive when the output plugs straight into revenue or compliance. Healthcare is a clean example, and not a counterexample: clinical documentation survives full pricing precisely because the practice or the payer absorbs the cost, not the clinician paying out of pocket for a novelty. That is the whole point. Value so obvious a business eats the cost is the surviving category. “Summarize my email for free forever” is the one I would bet against.
None of this means the technology was a bubble or that adoption reverses. It means the pricing was a promotional period, and promotional periods end. The interesting question for the next two years is not whether AI is useful. It clearly is. The question is which everyday habits people actually keep once they are paying what a query costs to serve, and which ones quietly go back to being a thing you used to do for free.