Claude Code's +50% Promo Was Extended Four Times

When this article went out on August 28, its headline said Anthropic “says this time it really ends on August 31”.

The next day Anthropic extended it again.

So this piece has to start with a correction: it didn’t end on August 31, the promotion ran all the way to September 13, and it was extended four times in total, not the three the original said. Now that the story is over, I’ve reorganized the whole sequence, because in hindsight those four extensions are a very clear signal in themselves, and what they speak to is whether AI companies have enough compute.

What was the promotion to begin with?

On May 13, Claude’s developer account posted this announcement:

Figure: ClaudeDevs announcing the +50% weekly limit on May 13 (source: x.com)

The content is simple. Claude Code’s weekly usage limit goes up 50%, through July 13, already applied, no need to opt in.

Notice that date: July 13. From the start it was a short two-month campaign.

How many times was it extended?

Four. I lined up the official posts I could find with the support center’s record:

Figure: the four extensions of Claude Code’s +50% weekly limit promotion (compiled for this article from ClaudeDevs posts on X and Anthropic’s support article)

The first extension was only six days, from July 13 to July 19. A delay of that length looks more like nothing had been decided internally and some time was being bought.

The second was a month, to August 19. The third went to August 31. The fourth went to September 13 and announced the ending at the same time.

One pattern is worth your attention. Every extension was announced a day or two before the expiry date: the extension to August 19 was announced on July 18, with one day left before July 19; the extension to August 31 was announced on August 18, again with one day left before August 19; the last was announced on August 29, with two days left before August 31.

For users, that means you could never plan ahead. You didn’t know whether next week’s allowance was 150 or 100 until the last moment.

How did Anthropic explain it?

The post for the third extension is the most candid in the whole thread:

Figure: ClaudeDevs announcing the third extension and mentioning tight capacity (source: x.com)

It says three things. Extended through August 31. We hope to make this a permanent change to our plans. But demand for the models is strong, and capacity may be tight over the coming weeks.

Read the second and third sentences together: they wanted to give it and couldn’t afford to.

I think that statement can be taken at face value, because it does Anthropic no favors. A company doesn’t say publicly that it’s short of capacity for no reason. And what it did afterwards matches: the permanent increase finally settled on is only 25%, half of what it wanted to give.

The support article, as later rewritten, is more direct in its opening note. It says the original plan was to return weekly limits to their original levels, but they know many people found the extra usage helpful, so while “the full promotional boost couldn’t last”, they’re making part of it permanent.

“Couldn’t last” is the answer to the four extensions.

What did the promotion actually raise?

This is where it’s easiest to go wrong. Many people took +50% to mean “you can use Claude fifty percent more”. The scope was much narrower.

Figure: what the +50% promotion raised and what it left alone (compiled for this article from Anthropic’s support article)

Three points.

Only the weekly limit, not the 5-hour limit. Claude Code has two layers of limits, one on how much you can use every 5 hours and one on how much you can use each week. The promotion touched only the latter. The support center’s wording is explicit:

Figure: the support article stating the promotion affected only weekly limits (source: support.claude.com)

What does that mean? If your pattern is intense use in short bursts, the layer you hit is the 5-hour one, and the promotion did nothing for you. It helped people who use it steadily every day and would hit the weekly limit as the week adds up.

An example. Suppose you use a steady 18% of the original weekly limit every day. Five working days come to 90%, just enough, and you hit the wall if you want to do a bit more at the weekend. With the increase, the same usage is only sixty percent of the new limit, and you gain a whole weekend of headroom. Conversely, if your habit is to pick one day a week and go hard, what stops you is that day’s 5-hour limit, and no increase to the weekly limit changes that.

Only Claude Code. Whether you use Claude Code from the command line, an IDE extension, the desktop app or the web, it counts. But limits for regular Claude chat, and for Cowork, didn’t change.

Not every plan. Pro, Max, Team and legacy seat-based Enterprise were eligible. Free plans and consumption-based Enterprise seats were not. The latter makes sense: you already pay for what you use, so there’s no “limit” to raise.

The support article also mentions a small tool: typing /usage in the CLI shows your current usage and limits. If you’ve never typed that command, you don’t actually know whether these four months of extra allowance made any difference to you.

Why call this a compute signal?

Looked at alone, you could call it a poorly managed marketing campaign. Put beside other things from the same period, it looks less like one.

In early September, a week after OpenAI’s GPT-6 Astra launched, the $200 ChatGPT Pro plan was pulled from sale outright. The stated reason was nearly the same sentence: demand is unprecedented and existing users’ experience comes first. Twenty days later it went back on sale at the same monthly price with about half the usage.

Two companies, the same month, the same shape: the allowance given to the heaviest users gets adjusted downward, and capacity is the reason both times.

Further upstream it’s the same story. In September there were reports that the three major memory makers’ 2027 capacity is already sold out. Whether model companies can buy the compute they want is no longer only a question of money.

My own view: subscription AI tools are going through a repricing, just not in the form of a price increase. The number on the price tag hasn’t moved; what moves is how much that number buys. Claude Code gave you 150 first and then told you it would be 125 from now on; ChatGPT Pro stopped selling and came back with half the amount. For a company, adjusting an allowance is easier to explain than adjusting a price.

What would change my mind? If any of them materially restores an allowance in the next six months, and not as another promotion with an end date. So far I haven’t seen it.

How did it end?

On August 29, Anthropic announced that from September 14 the weekly limit would be permanently 25% higher than before the promotion. Until then, the +50% stayed in effect.

So against the period before May, you have 25% more. Against the past four months, you have about 17% less. How both numbers hold at once, and how Anthropic put it, I’ve written up separately in the next piece.

There was a follow-up after that too. When Claude Opus 5.5 launched on September 23, Anthropic threw in a usage reset coupon with an expiry date. Softening a smaller allowance with a one-off compensation is a move we’ll probably see again.

What can you take from this?

If your workflow depends on a subscription tool’s allowance, I’d suggest three things.

First, record how much you actually use. Look at the numbers from commands like /usage once a week. Only by knowing what fraction of the limit you normally reach can you judge whether an adjustment affects you.

Second, keep “promotion” and “plan contents” separate. An increase with an end date isn’t something you bought, even if it’s extended four times. The support article’s terms are clear: the offer had no cash value, was not transferable, and could not be combined with other offers.

Third, have a fallback. Across these four months, each time you learned next week’s allowance only a day or two before expiry. If your work can’t tolerate that uncertainty, you should have another tool or another model to switch to. If the terminology around limits is unfamiliar, the glossary covers it.

What this rewrite corrected

• The original headline and body said August 31 was when “this time it really ends”. In fact it was extended the next day to September 13.
• The original said three extensions; the correct number is four.
• The original was written before the final extension and so had no ending. This version adds it.
• I confirmed the announcement dates of the extensions this time from the posts’ own timestamps, in US time. For the first extension (to July 19) I didn’t find the post for its exact announcement date, only media coverage.
• I have no way of knowing Anthropic’s actual capacity figures. “Capacity may be tight” is what the official post says, and my reading is built on that sentence.