At the end of September, OpenAI wrote to its Pro subscribers about some changes to their plans. By way of thanks, each existing subscriber would receive 62,500 usage credits, valued at $2,500 and expiring on 31 December. That’s four cents a credit, and the email is admirably precise about it.
The changes themselves got vaguer treatment. From 30 October, the $200 plan’s allowance for agentic work would fall from twenty times the Plus allowance to ten, at the same price. And how big is the Plus allowance? OpenAI publishes it as ranges of messages that vary with the model and the task, which is an honest way to describe something that can’t be pinned down.
I read the email twice, because it tells the whole story of AI pricing on a single page. When a vendor sells you more, it measures to the cent. When it describes what you have already paid for, it switches to multiples.
Elena Verna, who leads growth at Lovable, published a cheerfully self-incriminating teardown of AI pricing pages in mid-October. She asks of one vendor after another what “5x more usage” is five times of, and admits that her own company sells credits customers can only value once they start building. I recommend it. Her diagnosis is confusion. I think the confusion persists because it does a job for the vendor, and that there is a point where an experienced buyer should stop putting up with it.
My position, briefly: for small use cases, vague allowances are a fair trade. For larger, mission-critical ones, you cannot accept vagueness on something that, if changed, could drastically affect budgets or outcomes.
Focus On: Priced on the Way Out, Vague on the Way In
OpenAI isn’t unusual. The largest vendors describe the two halves of the same subscription in different currencies.
Anthropic sells its Max plans as five or twenty times Pro. Once you exceed them, extra usage is billed at standard API rates, per million tokens. SpaceXSI’s new X Subscription lists its $8 Plus tier simply as the base, with Premium, Super and Ultra at four, fifteen and fifty times that. Third-party models used through Cursor, which SpaceXSI now owns, are billed at API list price through top-ups. Google moved Gemini to compute-aware limits in May that vary with the complexity of your prompts, while selling AI credits as a currency for its video and agent tools. OpenAI has priced Codex in credits per million tokens, at its API rates, since April.
Each of these companies has a perfectly good unit and uses it to bill extra usage. For the allowance you have already paid for, you get a multiple, an adjective, or a meter that fills at a rate you can’t see.
GitHub is the exception. In June it moved every Copilot plan to usage-based billing, which on its own sounds like the opposite of what buyers want. But it wrote the allowance in the same unit as the overage: the $10 plan includes $10 of AI credits, the $39 plan includes $39, and a credit is worth a cent. Because both sides were already in money, GitHub could show customers a preview of their projected bill a month before the switch.
Why does this matter more than a confusing pricing page? Because an allowance described as a multiple can be resized without anything that looks like a price change. Since OpenAI’s change, every Pro tier costs the same $20 per Plus unit, where the $200 plan used to cost $10. A Pro subscriber’s invoice doesn’t move, and the agentic work it buys halves.
The pattern belongs to no single vendor. Anthropic made Claude’s five-hour sessions run down faster during weekday peak hours in March, then doubled Claude Code’s five-hour limits in May. Google tightened Gemini’s limits after its developer conference and tripled them within a day of the complaints. Allowances move in both directions, which is rather the point: the decision sits with one party. SemiAnalysis, which now measures these meters for a living, found one of its test accounts with roughly a fifth less allowance than identical ones, and the provider confirmed a very small experiment on limits.
Whatever the intent, this works like an option. The vendor holds the right to resize what you bought, and the customer has written that option, usually without noticing.
Why the Option Exists
Many subscribers have read all this as cynicism. Mostly, I don’t think it is.
Gartner expects the inference cost of an agentic workflow to rise more than fivefold through 2028, even as the price of each token keeps falling. Agents plan, call tools, retry, and re-read their own context, so the bill follows the work rather than the token. SemiAnalysis estimated in June that a fully used $200 plan could consume between $8,000 and $14,000 of tokens a month at API prices, depending on the vendor. Those are API-equivalent values, not the vendors’ costs, but the direction is clear enough: flat subscriptions are subsidised, and heavy users get most of the subsidy.
This is the best argument for vagueness. An undefined allowance lets a vendor hold the entry price and be generous to the people who push hardest, while keeping room to adjust as its costs move. If the option has a premium, the subsidy is it: heavy users are, in effect, paid to write it. Define the unit and the generosity becomes a visible number, and so does any reduction. Many heavy users would probably pay more under a fully transparent model.
Verna adds a practical defence of credits themselves. A product that routes work across several models and tools needs a common currency, or customers end up managing a miniature cloud bill for every task. I agree. The difficulty is the exchange rate, which vendors publish for top-ups and leave blank for the allowance.
So I accept the bargain for small use cases. If a team is experimenting, or a tool supports work that could pause for a week without anyone outside the team noticing, the subsidy is worth more than the certainty.
Where Vagueness Stops Being Acceptable
For now, the line is spend. Most organisations already know which AI tools carry real budget, and those deserve an allowance written in a unit Finance recognises, not a multiple of a plan nobody has defined.
The line will move once agents run critical workflows end to end (campaign operations, say, or the reporting that feeds the board). At that point the question stops being what an overrun costs and becomes what stops working. An allowance halved at renewal is a budget problem; an allowance exhausted halfway through a month-end close is an outage, caused by someone else’s pricing decision.
Picture an agent that assembles the monthly performance pack from a dozen data sources. In a quiet week it fits comfortably within the plan. In the last days of a quarter, when every team wants its own cut of the numbers, it can burn through a week’s allowance in an afternoon. On a plan described in multiples, nobody can tell in advance whether the allowance will hold.
That changes what I would require before anything material goes live: a clear plan B for the day you can no longer commit to the spend. This is ordinary business continuity planning (the supplier just happens to describe its capacity in adjectives), and the right plan B depends on the work. Where volume is high and any capable model could do the task, it is a second model or vendor, tested in advance rather than in a crisis; the portability case I made in September applies directly. Low-volume work that depends on judgement needs a human fallback with named owners. And high-volume work that can tolerate some loss of quality needs a degraded mode agreed beforehand, such as a cheaper model or a narrower scope.
What to Do This Quarter
If you sell AI, the fix costs very little, because the unit already sits on your overage page. Quote the included allowance in that same unit, with two or three representative tasks and honest ranges. Make soft caps and early alerts the default: in a survey of 296 software buyers by Pricing I/O and Benchmarkit, 62% wanted soft caps with approval to continue, and pre-purchased credit pools came last. And preview the bill before you change it, as GitHub did.
If you buy AI, ask one question at every renewal of anything material: what is my allowance, in the unit you would bill me for extra usage? A vendor that can’t answer has told you the plan B is yours to build. A vendor that can should have little difficulty writing the rest into the contract. For the tools that matter, I would ask for four things:
the allowance in that billing unit, with a few representative tasks attached;
notice before any reduction, on the terms you would expect for a price rise;
the right to leave, or a pro-rata credit, if the allowance falls mid-term;
a usage export Finance can read without asking the vendor.
The vendor already measures every one of those numbers.
Buyers shouldn’t feel too righteous, though. In the same survey, nearly nine in ten had exceeded their initial AI budget, yet only one in ten blamed vendors for changing prices or terms after the sale. Most pointed to their own forecasting: 67% said AI features and automation had driven more usage than planned. So part of the fix sits inside our own organisations. Put Finance next to IT on the AI meter, since IT owns AI cost risk in two-thirds of companies and Finance in fewer than one in five. And give your agents the stopping rules that keep consumption tied to value.
Where This Lands by the End of 2027
My bet is that large buyers fix this before pricing pages do. The organisations negotiating bundles and custom deals have the bargaining power to get the allowance written into the contract, in money or in API-rate credits. Some already buy that way: OpenAI’s enterprise customers on flexible pricing work from credits rather than fixed limits. By the end of 2027 I expect a two-tier market, in which the largest buyers know exactly what they have bought and everyone else is still buying multiples.
Regulators will reach the long tail, more slowly. The UK’s new rules on subscription contracts start in January, and the EU’s Digital Fairness Act is expected in draft on 11 November. Neither was written with AI allowances in mind. Still, European consumer authorities have already said that games must show the real-money price of virtual currencies. In the UK, the test for a misleading omission asks what the average consumer needs to take an informed decision. Sooner or later someone will put that question to an AI pricing page. I would rather vendors answered it first.
As for the shape of pricing itself, the buyers in that survey ranked pure outcome-based pricing last of seven models. Vendors in a separate survey of 230 software companies, by contrast, expect its share to grow from 5% to 31%. Buyers keep asking for a fixed anchor and a meter they can read. I suspect they’ll win that argument, at least until 2027 is out.
Before your next AI renewal, find your most critical workflow and ask which side of that two-tier market it sits on. If the honest answer is “multiples”, you have your first action for the quarter.
Follow me
That’s all for this week. To keep up with the latest in generative AI and its relevance to your digital transformation programmes, follow me on LinkedIn or subscribe to this newsletter.
Disclaimer: The views and opinions expressed in Chronicles of Change and on my social media accounts are my own and do not necessarily reflect the official policy or position of S&P Global.
