Why your AI subscription keeps changing its limits
A fixed monthly price sits on top of a variable meter nobody publishes. That gap explains every AI limit change of 2026, including the next one.
Here is a question you probably cannot answer about the AI subscription you pay for. How much can you use it this month?
Go and look. The plan page offers "more usage", or 4x, or 20x. The help centre says limits apply and depend on what you ask for. Settings may show a percentage bar, which tells you how much of something you have spent without ever naming the something. Underneath all of it sits a real number in a real unit, and it belongs to the company rather than to you.
That gap is the whole story. It explains why your limits keep moving, and why the same argument broke out about Claude in August and about Gemini in May. It will explain the next one too, whichever provider it happens to.
The meter behind the price tag
A subscription is a fixed price for a variable amount of work. Every message you send costs the provider an amount that swings wildly: a two-word question is nearly free, a forty-page PDF plus six follow-ups is not, and a video generation sits in a different bracket again. Your payment does not move. The cost of serving you moves constantly. The usage limit is what absorbs the difference, and it is the dial a provider turns when demand rises or capacity arrives late.
The companies are fairly open about this once you read past the marketing. When Anthropic doubled Claude Code's five-hour limits for good on 6 May 2026, it put the news in the same post as a deal for more than 300 megawatts of new capacity. When it took weekly capacity back four months later, the announcement closed with "thanks for hanging with us while we figured out what we can sustainably serve going forward". Google's help centre is flatter still, and says in one line that "Limits may change without notice, including due to capacity constraints".
None of that is scandalous. Capacity really is scarce and really does move. The trouble starts one step later, when the company has to describe the dial to the person paying for it.
Why every unit fails
Providers have tried four units, and each breaks differently.
Messages are what people naturally count, which works right up until you notice that no two messages cost the same. Say a plan caps you at forty a day. Generous if you ask quick questions, miserly if you feed in documents, and the company gets blamed for a number that looked fair on paper. OpenAI's answer was to stop counting: on 6 August 2026 it announced that text message limits were coming off every ChatGPT tier, free accounts included. Abolishing the unit solves it for text and leaves it standing for files, images, voice and image generation.
Hours were Anthropic's choice when weekly limits first appeared in July 2025, with Pro published as 40 to 80 weekly Sonnet hours. Nobody experiences an hour of a model, though. It is a proxy for compute wearing friendlier clothes, and it gives you nothing to plan a week around. Anthropic has not republished those figures since, so every "Claude Pro gets X" claim circulating today traces back to one thirteen-month-old announcement.
Tokens are precise, auditable and useless to a general audience. Compute is the most honest of the four and the least legible, which is what Google discovered in May.
So providers reach for a fifth option. They sell you a multiple. Two times, four times, twenty times. A multiple needs a baseline to mean anything, and the baseline is exactly the figure nobody publishes, which is how a plan page makes a precise-sounding promise about an unknown quantity. A Hacker News commenter posting as partsch, reading the late-August announcement, put the consumer-protection version of it better than any analyst has:
How can it be legal to offer a service without specifying exactly what you're getting?
A promotion becomes the number you feel you own
The Claude episode is the cleanest worked example, because nothing about the underlying plan actually got worse.
A temporary 50% boost has sat on top of Claude Code's weekly limits since 13 May 2026, extended again and again across the summer. On 29 August, Anthropic said that from 14 September it would permanently raise the standard weekly limit by 25%, then added, one post later, that measured against today the same change is a 17% cut. Both figures are correct. Measured from the contractual base, subscribers end September with more than they had in April. Measured from Tuesday, they end it with less, and Tuesday is the only baseline anybody actually uses.
The mechanism is general. A promotion that runs four months and gets extended five times stops being a promotion in the mind of the person receiving it. It becomes the size of your week, the amount of work you assume you can finish by Friday. Withdrawing it is arithmetically an increase and emotionally a confiscation, and the company collects the backlash for the second reading however carefully it words the first.
Google's May experiment in honest metering
Google took the opposite route and got a version of the same trouble.
On 20 May 2026, after I/O, the Gemini app switched from message counts to compute-based limits. How much you get through now turns on how hard a prompt is to answer, which features it pulls in, and how far the conversation has already run. As a description of what costs the company money, that is far more truthful than counting messages. As something you can plan around, it is worse, because you cannot know what a prompt costs until you have spent it.
The complaints arrived within a day, and the sharpest was about the baseline again. Google's multipliers for AI Plus and AI Pro are measured against the free tier, so a Pro subscriber reading "4x" learned how they compared with an account paying nothing, and nothing at all about what 20 May had done to their own.
What Google did next is why this example belongs beside the Claude one. Within eight days it shipped fixes, and almost none of them were a bigger number. It capped how much of your quota a single prompt could consume. It stopped charging users for its own failed requests. The same thinking carried on past May: from 2 September the metering reaches Gemini Notebook, carrying a cost indicator that shows what a generation will consume before you run it, and a queue for the heavy jobs.
That is what taking the meter seriously looks like. None of it made the allowance bigger, and all of it made the allowance steerable, which is worth more to a subscriber than another multiplier. The size of the budget remains Google's private information.
What good disclosure would actually look like
Set the bar low and see who clears it.
Publish a figure, in a unit, with a date on it, even as a range. Anthropic managed that once, in July 2025, and has published nothing numeric since. Name the baseline wherever the multiplier appears, on the same page, in the same size type. If a reader has to open a help centre article to decode a pricing headline, the headline is doing work it did not earn. Label a promotion as a promotion, with its end date, on the day it starts. Show cost inside the product before the spend, which Google has now proved is possible. Announce reductions with the same reach as increases. All five are cheap, and nobody is doing all five.
Read the plan page like a contract
We move our own limits, so treat this as a disclosure. Free accounts got web search for the first time in April 2026, with a monthly allowance attached from the first day, later raised to two. Until the end of June the button in our web app still said one free search a month while the server was granting two: a visible number and an enforced number that were not the same, which is the small version of the sin described above. The label now derives from the value the server checks, so the two cannot drift again. In August we reworked free uploads: three a day where it had been two, the monthly cap gone, 10MB a file where it had been 5MB.
Those moved one way, and others will move the other way in time, because we face the same arithmetic as everyone else. So here are ours, in units: 50 messages every 24 hours on the free plan, two image generations and two web searches per 30 days, three file uploads a day at 10MB each. Our own limits page prints one of those four; the others show up as "limited" or not at all, which fails the first test on the list above. Writing them here and not there is the wrong way round. Paid plans sit in the pricing section.
Hold anyone to the same standard, and add the question the plan pages almost never answer: what happens when you actually hit the limit? A wait until 3pm and a hard stop until Monday are very different products sold under the same word.
The multiple is the least informative thing on the page. What deserves your attention is whether the company will tell you, in advance and in writing, what you are about to spend, and whether it tells you again when that changes. Everything else in AI pricing will keep moving, and it should. The unit is the part that has no excuse.