ATAILA Newsroom · Budapest · 2026-09-05

2026-W36

On 13 September our Claude Code allowance drops by a third. That is the trend, not the exception.

Anthropic's own support page states it without decoration: "From May 13, 2026 through September 13, 2026, your weekly usage limit in Claude Code is 50% higher." And then: "After September 13, 2026, weekly usage limits in Claude Code return to their standard levels." We are a paying Claude Code customer. In eight days the ceiling we have been building against comes down by a third — not because anyone broke a promise, but because the promise had an end date. This is worth writing about precisely because nothing went wrong.

What the page actually says

No interpretation needed — the numbers are published.

That last figure matters, and it is the reason this is not a complaint. A bonus is expiring on the date it said it would, after having been extended once already. Nobody raised a price. Nobody cut anyone below the baseline they bought. If you want a story about a vendor behaving badly, this is not it.

It is still true that the same subscription buys less work next week than it bought last week. Both of those things are the case at once, and only one of them shows up in a budget.

The same week, the same direction, a different vendor

In April, GitHub announced that Copilot would move from flat premium-request units to usage-based billing on token consumption, effective 1 June. The reasoning was stated plainly, and it is the most quotable sentence any AI vendor has published this year:

> "GitHub has absorbed much of the escalating inference cost behind that usage, but the current premium request model is no longer sustainable."

Read that as what it is: a vendor confirming on the record that it had been subsidising the product, and that it has stopped. Input, output and cached tokens, at each model's published API rates, across Pro, Pro+, Business and Enterprise.

Two of the largest AI coding vendors in the world, one quarter, the same move — one by ending a bonus, one by changing the billing model. We wrote separately about the third signal, Google shipping a model that costs the same per token and up to 40% more per task.

Why this keeps happening, and why it will continue

It is tempting to read these as three unrelated commercial decisions by three companies having an awkward quarter. They are not, and the arithmetic is public.

PwC's datacentre outlook puts roughly $800 billion of datacentre investment in 2026 alone, rising toward $1.8 trillion a year by 2050. Set against that the detail that decides everything: the useful life of the GPUs at the centre of it is four to six years. Then add what the buildout costs to merely keep running — power at industrial scale, and cooling, which is now a siting constraint rather than a line item.

Capital on that scale, depreciating that fast, with an operating cost that never sleeps, has to be recovered from someone. The someone is whoever is renting the inference. The price that was set to win a market is not the price that services the balance sheet behind it, and the gap between those two numbers is closing in one direction only.

That is the part worth planning around. Not a grievance about any one vendor — a reading of where the money has to come from.

What we think — including the part that cuts against us

Here is where we could overclaim, so we will not. The tempting pitch is buy your own hardware and you only pay for electricity. Taken literally that is false, and our own private AI page says so. Buying a box gets you a box. The patching, the backups, the GPU that fails at three in the morning, the model lifecycle, the depreciation — all still yours, and all of it is the expensive part.

The accurate version is narrower and, we think, more useful:

If you own the hardware, you pay the electricity — and we bring the rest

That is the actual shape of the offer. Your marginal cost of the next inference really is close to power, because the machine is yours and it is already paid for. What stands between "we own GPUs" and "a confidential workflow runs in production with monitoring, backups, a release path and an SLA" is a platform, and building that yourself is the project that eats the savings. That platform is what we operate. You are not buying compute from us; you are buying everything around it.

It changes what your cost depends on

On metered public AI your run-rate is a function of another company's pricing decisions, release schedule and funding pressure. On dedicated capacity — ours or yours — it is a function of your own workload. It is not automatically cheaper on day one. It is forecastable, and it does not move because a vendor's board decided the subsidy phase is over.

Our own relevance goes up when public AI gets dearer, which is exactly why we should be careful

We have an obvious interest in this trend continuing, so we would rather be caught understating it. We use public models ourselves where the data allows, including Claude, and we will keep doing so. This is not a call to move everything in-house.

Where we stand

We are not annoyed at Anthropic. They ran a promotion, extended it, and ended it on the published date — that is what a promotion is, and being able to quote the page is a mark in their favour, not against it.

The point is the direction. In a single quarter, two of the most-used AI development tools in the world have either started charging for what used to be included, or stopped including as much. The capital behind the GPU buildout says that continues. If your plan for next year assumes AI gets cheaper per unit of work, that plan needs a second column — and if a workflow matters enough to run every day, you should be able to say what it will cost in twelve months. On metered public AI, right now, nobody can honestly tell you that.