Anthropic Announced a 25% Raise. Then Deleted the Post.

On August 29, 2026, Anthropic’s developer account announced that Claude Code weekly limits would rise permanently by 25% from September 14 for Pro, Max, Team plus seat-based Enterprise plans. Developers worked out within minutes that the temporary 50% boost expires the same day, so the real change is 150 down to 125, a 17% reduction against current allowances. The backlash was strong enough that Anthropic deleted the original thread and republished it, this time stating the 17% cut plainly. A named Claude Code power user announced publicly that this is his last week using it as a daily driver. All of it landed in the same fortnight that Sonnet 5 repriced 50%, with OpenAI restoring the Codex five-hour limit. Best for anyone budgeting agent work. Not ideal for anyone who took the headline at face value.

The post said limits were going up 25%.

Developers checked the calendar. The temporary 50% boost, which has been running since May, expires on the same day the permanent raise begins.

So if your baseline was 100, you have 150 today. On September 14 you get 125.

That is a 17% cut, announced as a 25% increase.

The arithmetic took the developer community about ten minutes. Anthropic deleted the thread and posted it again, this time saying the quiet part in the first line.

What Was Actually Announced

Verified across nine sources including BleepingComputer, Android Headlines, Digital Applied plus explainx.

The republished wording is the notable part. Anthropic now says the change works out to a 17% reduction in weekly limits compared to today, then adds that it is working on changes making it feel like you are getting more from Claude, with more visibility plus control over usage.

That second sentence is doing a lot of work. Feel is not a unit.

The Reversal Is Not Symmetric

Here is the detail almost nobody has run. It matters for anyone planning a budget.

Going from 150 to 125 is a 16.7% decrease, measured against 150.

Getting back from 125 to 150 requires a 20% increase, measured against 125.

So a future announcement restoring a 17% rise would not put you back where you started. Percentage direction is not reversible. Any headline promising a percentage without naming the baseline is telling you almost nothing.

That is not sleight of hand on Anthropic’s part. It is ordinary arithmetic that ordinary announcements routinely obscure, which is exactly why the developer community re-derived it in minutes rather than accepting the framing.

Digital Applied ran the same calculation off Anthropic’s own three numbers and reached the same place.

The Documentation Was Wrong While the Announcement Ran

A second thing worth knowing, since it says something about how these changes reach customers.

When Digital Applied retrieved Anthropic’s own promotion page on August 30 at roughly 09:58 UTC, a full day after the announcement, it still described the boost as ending August 31. It was still titled as the May to August promotion. It still said terms were valid through August 31 at 11:59 PM PT. No mention of September 14. No mention of the permanent increase.

Its FAQ answer for what happens next, saying weekly limits return to standard levels after August 31, had been superseded twice by then. The boost now runs to September 13, while the standard level itself is changing.

So for at least a day, a customer reading the official documentation would have got a different answer from a customer reading the announcement. Both were current.

That is the practical problem with metered products where the meter is a policy. The number changes faster than the page describing it.

A Power User Quit in Public

The reaction is where this stops being a pricing note.

Kieran Klaassen, who builds Cora and has been one of the more visible Claude Code users, posted that this is the last week he uses Claude Code as his daily driver. That post has drawn over 156,000 views plus 907 bookmarks.

Bookmarks are the signal there. People are not just reading it. They are saving it, which usually means they are considering the same move.

Theo, who runs one of the larger developer channels, posted about the change and noted that Anthropic reached out to clarify the 50% boost was always temporary while the 25% is permanent. Both statements are true. Neither addresses that the meter still goes down on September 14.

Matt Shumer ran the numbers publicly too.

Anthropic’s Claude Code lead responded in the threads, which is more engagement than most vendors offer. It did not change the arithmetic.

The Boost Was Always Going to End

Being precise about what Anthropic actually promised, because the defence has some merit.

The 50% weekly limit boost started in May 2026 and was always described as temporary. It has been extended repeatedly. It was due to end July, then August 31, now September 13. Each extension came with language about hoping to make higher limits permanent.

So a customer paying attention knew the boost was borrowed capacity. Anthropic’s clarification to Theo, that the 50% was always temporary while the 25% is permanent, is accurate.

The problem is what four months of extensions does to expectations. A temporary measure renewed three times stops feeling temporary. People planned around it, built workflows that fit inside it, then discovered the ceiling was coming down.

That is a communication failure rather than a broken promise. Nobody was lied to. Plenty of people were surprised, which for a paid product is its own kind of cost.

And it points at something structural. Every one of these boosts, extensions plus reversals happened through posts on a developer account rather than through account emails or the billing page. If your quota is a policy announced on social media, the only people who track it reliably are the ones already following the account.

What the Reaction Actually Measures

Worth separating the noise from the signal, since a backlash is easy to overstate.

The noise is the usual complaint volume any pricing change generates. That tells you nothing.

The signal is in two places. First, the bookmark counts. Kieran Klaassen’s post about leaving drew 907 bookmarks against 1,637 likes, which is an unusually high ratio. Likes are agreement. Bookmarks are intent to return, which for a post about switching tools means people are treating it as a decision to make rather than a sentiment to endorse.

Second, the deletion itself. Companies do not remove announcements over ordinary grumbling. Removing a post and rewriting it with the unflattering number in the first line is a measurable admission that the original framing failed.

What that does not tell you is whether anybody actually leaves. Switching costs for a coding agent are real: muscle memory, configuration, project context, plus whatever the workflow has grown around it. Announcing a departure on X is considerably cheaper than completing one.

The number worth watching is not this week’s reaction. It is whether the Codex bridging tools we tracked in yesterday’s roundup keep gaining forks through September.

This Is the Fourth Change in a Fortnight

Line up the last three weeks and the pattern is hard to miss.

On August 26, OpenAI restored the five-hour rolling limit on Codex for Plus subscribers, six weeks after removing it. Some developers had switched to Codex specifically because it was gone.

Three days later Anthropic announced the Claude Code change, deleted it, republished.

Today, Sonnet 5 introductory pricing ended. Input went from $2 to $3 per million tokens, output from $10 to $15.

September 14, the weekly limits drop 17%.

Meanwhile DeepSeek scrapped flat API pricing for peak and off-peak rates that run higher even at the discount, plus raised V4 Flash by 93%.

Every major vendor tightened something inside one fortnight. None of them coordinated. They are all responding to the same thing, which is that agent workloads consume compute in a way flat subscriptions were never designed to absorb.

What Anthropic Is Actually Dealing With

Being fair to them, because the constraint is real.

A chatbot conversation costs a predictable, small amount. An agent session running for hours with tool calls, subagents plus repeated context consumes orders of magnitude more. Consumption also varies enormously between two customers paying the same twenty dollars.

Flat pricing assumes rough parity. Agents destroyed that assumption. One Claude Code user might consume a hundred times another on the same plan.

Anthropic also just posted its first profitable quarter on revenue above $11.5 billion, driven substantially by enterprise Claude Code adoption, while sitting weeks from a public listing. Private companies can price for growth. Public ones price for margins, so quota is where that gets expressed when the headline subscription price stays the same.

None of which makes the framing better. It explains why the pressure exists, not why the announcement led with the flattering number.

The Trust Problem Is the Real Story

There is context here that makes the deletion land harder than it otherwise would.

Yesterday we covered the March cache bug, where a string replacement in Anthropic’s custom runtime broke prompt caching, charging some subscribers ten to twenty times more than they should have paid for about a month. Customers found it by reverse-engineering the binary. Separately, Anthropic had cut the default prompt cache lifetime from an hour to five minutes with no announcement at all.

There is also an existing Claude Max usage-limits lawsuit where subscribers contest whether marketed multipliers match delivered capacity.

So the sequence over six months reads: limits get silently tightened, a bug inflates consumption by an order of magnitude with no notice, customers diagnose both themselves, then a limits reduction gets announced as an increase and pulled after the community objects.

Each of those has an innocent explanation. Together they describe a product where the customer cannot verify what they are buying, then repeatedly turns out to be the one who discovers the number moved.

/separator

What You Should Actually Do

Practical, since September 14 is two weeks out.

Find out what you use now. Claude Code writes session logs to ~/.claude/projects/ recording tokens per call. The tools that read those logs will tell you your real consumption by model, project plus day. Do it while the 50% boost is still active so you know what a full week actually costs you.

Write the number down outside the tool. Every one of those trackers prices from a hand-maintained table. Sonnet 5 changed today. They will under-report until somebody ships a patch.

Work out whether 125 covers you. If your current usage sits under two thirds of the boosted allowance, nothing changes for you. If you regularly hit the wall at 150, September 14 makes that worse rather than better.

Then decide

Look at where the tokens go. Output costs six to ten times more than input. A workflow producing verbose output costs far more than one reading widely then replying briefly. That split is usually the cheapest thing to fix.

Check what the boost is worth to you specifically. Between now and September 13 you are running at 150. If you never come close to the ceiling, none of this touches you and the permanent 25% is a straight improvement over your real baseline. The cut only bites people who were already using the extra headroom, which is a smaller group than the reaction suggests.

Know the alternatives before you need them. Codex has its own five-hour window again. Local inference has no quota at all. We covered an engine running a 35B model at 39.3 tokens per second on an 8GB laptop GPU, which beats the median decode speed in real Codex production traces. That is a hardware purchase with a payback period rather than a research demo.

Our Claude Pro review covers what each tier includes at today’s prices.

What Would Have Made This Fine

Worth saying, since the underlying decision is defensible while the handling was not.

A post that opened with the actual net change would have cost Anthropic one uncomfortable news cycle and bought considerable credit. Something like: the temporary boost ends September 13, standard limits rise 25% permanently, so heavy users see roughly 17% less than today, then here is why.

Every fact in that sentence is one Anthropic published. The only difference is which number goes first.

They arrived there eventually, after deleting the original. The republished post says the 17% cut plainly, which is the right correction. It just came after the community had already done the work, which is a materially different thing from leading with it.

The line about upcoming changes making it feel like you are getting more from Claude is the part that will get quoted back. Customers who just recalculated their own quota are not the audience for feel.

There is a version of that sentence that would land, incidentally. More visibility and control over usage is exactly what people have been asking for through every incident this year. If what ships in September is a real usage dashboard showing tokens against quota in something other than a percentage bar, that is worth more than the 17%. The framing failed because the promise arrived attached to a reduction rather than on its own.

The Part Worth Keeping

An announcement said 25% up. The actual number was 17% down.

Both figures were true. One described the change against a baseline nobody currently has. The other described it against what people are using today.

Developers found the gap in about ten minutes, because the people paying for a metered product are the only ones motivated to check the meter.

Then the post came down and went back up with the honest number in it.

That is the system working, in the sense that the correction happened. It is also the fourth time this year that Claude Code users learned something about their own usage by working it out themselves.

September 14 is two weeks away. The logs are on your machine now.