Man, who can handle this?
On August 24, Claude crashed three times in a row.
Anthropic’s status page flashed three bright red alerts that day, and the official team repeatedly stepped in to announce that service had been fully restored.
But every time the page just turns green, there’s still an outcry on X, with many users complaining they still can’t access it.
This is the 13th day in August that Anthropic has recorded an incident.
Wow, only 24 days have passed this month.
You tell me that Claude spent more than half of his days lying in bed.
A Reddit user even complained that his automated process, which had run diligently for nine hours, was completely wiped out by this major outage, rendering all his effort useless.
529, the kitchen caught fire directly.
When the system failure fully erupted, users saw "529 Overloaded" pop up on their screens.
Anyone who has used an API is familiar with 429, which is rate limiting—similar to a restaurant telling you to slow down your orders and wait a moment.
And 529 represents the entire system being completely overloaded, equivalent to the restaurant staff running out to tell you, "Sorry, our kitchen is on fire."
According to the status page, the fire burned from 12:50 PM to 3:36 PM Beijing Time, lasting nearly three hours.
Twenty-one minutes after the alert was triggered, the official team remained confident, stating that the root cause had already been identified.
Then what? That’s it—there’s nothing more. Anthropic has not disclosed any technical reasons or provided a timeline for full restoration.
The list of affected victims is very long.
Mythos 5, Fable 5, Opus 5, and Opus 4.8 are all down; claude.ai, API, Claude Code, and Cowork are all experiencing critical issues.
In the end, only the Console and Claude for Government locations narrowly escaped destruction.
The entire Claude tech stack has basically been compromised.
Repeated sit-ups, turned red again, then green again, then red.
Even more absurd scenarios are yet to come. Every time the official team announces that repairs are complete, it promptly dies again.
The model's widespread errors lasted nearly three hours before being resolved.
The system crashed again just after midnight on the 25th, affecting all Claude.ai and Claude Code subscriptions, and it took six minutes to recover.
Not even a few hours of peace, and at exactly 4 a.m., it went down again—this time stuck for eight minutes.
A monitoring bot named isclaudedownbot, dedicated to tracking Claude's login status, tweeted at 4:11 AM—its entire message contained no fluff, just one word.
Yes.
You have to admit, this emotionless announcement was incredibly precise.
Anthropic’s official statement was as polished as ever, stating that services have been fully restored on Claude.ai, Claude Code, and the Claude API. The company acknowledged how much users rely on Claude and thanked everyone for their patience while the team worked to resolve the issue.
That said, a green status page does not mean your account is operational.
The tweet posted by netizen elstar that day no longer sounded like a complaint—it sounded more like a cry for help.
In reality, the official status currently shows everything as normal, but lingering authentication and frontend issues are causing some unlucky users to see a glaring white screen on their Settings, Usage, and Skills pages.
Everything is normal and the page is blank—two contradictory things that are simultaneously true right now.
Even late into the night, someone is still yelling online that Claude is down again.
The system is experiencing a malfunction.
And on this day, you're really not alone at all.
Third-party monitoring service StatusGator shows that, since January 2026, Claude has accumulated 184 outage records.
Keep in mind that it's only August right now.
And scrolling back through the status page, the entire month of August was filled with outages, dense and relentless like a calendar.
Two incidents occurred on the 20th; one each day from the 19th to the 16th, consistently. On the 15th, Fable 5 crashed for approximately four hours. On the 14th, it went completely off the rails, with three incidents in a single day. On the 13th and 12th, one incident each occurred without surprise.
The most devastating occurred on August 5, with a collapse lasting a grueling 7.5 hours.
The error message the user saw on the screen was: "Due to unexpected capacity constraints," which translates to "Unexpected capacity constraints."
Hash rate depletion—who moved my reasoning intelligence?
capacity constraints appeared more than once in the error reports this month.
On X, another phrase that is repeatedly mentioned is "ran out of compute," meaning exhausted computational power.
Some developers have been blunt: Opus 4.6 was once so smart it gave people chills, but then they seemingly ran out of computing power and weakened everything.
Netizen Eason also feels that Claude has experienced an unprecedented severe decline in intelligence; Opus 5 is completely unusable, and he has even reverted to using Opus 4.8.
Reverting versions proactively is truly rare in the hyper-competitive AI industry.
Of course, these are merely users' subjective feelings, lacking concrete evidence.
So someone brought the two together and ran a trade.
For the same challenging task, Opus 5’s xHigh tier takes a full hour, while switching to another provider’s equivalent model completes it in just 15 minutes.
A fourfold difference that leaves you breathless.
Even AMD’s AI Director, Stella Laurenzo, personally dug through her team’s entire archive of 6,852 Claude Code sessions.
When I pulled up the results, the curve showed a sudden cliff-like drop.
At the end of January, the model's thinking depth was approximately 2,200 characters; by early March, it had plummeted to a mere 560, a 75% decline.
An even more chilling coincidence is that during the same period, Claude Code began quietly hiding its thought process from users, and within just one week, the hiding rate increased from 1.5% to 100%.
In other words, as the model visibly becomes dumber, you can't even see how it's becoming dumber.
The agent worked for nine hours, but the status page recorded only a few minutes.
Anthropic lists the following 90-day availability rates on its official website: claude.ai at 99.33%, API at 99.43%, and Claude Code at 99.35%.
At first glance, it looks like a perfect score. But for enterprise-grade services, the passing line is 99.9%.
Moreover, in the era of the classical internet when the metric of availability was first introduced, service outages merely meant a webpage wouldn’t load—you could simply refresh it and get back to work, losing only a few seconds in the process.
However, for an agent, when a long task is interrupted, the cost is never just the few minutes lost during the break—it also includes all the time, context, and computational resources already spent. This ledger cannot be tracked, nor can it be tracked, by a status page.
Anthropic still has not explained what happened during those three hours on August 24.
And this August, which has been red for 13 days, still has seven days left.
Reference materials:
https://x.com/ns123abc/status/2091784366519193852
This article is from the WeChat public account "New Intelligence Yuan," author: ASI Revelation, editor: Moses