Anthropic’s Claude went down three times on Aug. 24, with outages affecting a wide stretch of its product stack, including claude.ai, the API, Claude Code and Cowork, according to the source article. Anthropic’s status page turned red three times that day, and the company also posted three separate notices saying services had been fully restored. Even so, users on X kept saying they still could not access the service from their side.
The article described Aug. 24 as Anthropic’s 13th incident day in August, with only 24 days of the month having passed at that point. One Reddit user said an automated workflow that had been running for nine hours was cut off during the outage, wiping out all of that work.
Users were met with 529 Overloaded errors
When the disruption spread across the platform, many users saw a 529 Overloaded message on screen. The article contrasted that code with the more familiar 429 rate-limit response used by APIs. In the framing used by the report, 429 means requests are being throttled, while 529 signals a full system overload.
Based on the status page record cited in the article, the main incident lasted from 12:50 p.m. to 3:36 p.m. Beijing time on Aug. 24, or close to three hours. Anthropic said 21 minutes after the alert that it had identified the root cause. It did not, however, disclose the technical reason or provide a full recovery timeline.
The list of affected components was broad. The source article named Mythos 5, Fable 5, Opus 5 and Opus 4.8 among the impacted models. It also said claude.ai, the API, Claude Code and Cowork were all affected. Console and Claude for Government were described as the only parts that escaped the incident.
Recovery notices did not end the disruptions
The report said the instability did not stop after the first all-clear. After the model-related errors ran for nearly three hours and were marked as resolved, the login system failed again just after midnight on Aug. 25, hitting claude.ai and Claude Code subscriptions. That disruption lasted six minutes. A few hours later, at 4:00 a.m., another login outage hit and lasted eight minutes.
An automated account that tracks Claude login status, isclaudedownbot, posted at 4:11 a.m. with a single word: 「Yes.」 The article presented that as a blunt summary of the moment.
Anthropic’s official notice said service had been fully restored across Claude.ai, Claude Code and the Claude API, and that the company understood how much users rely on Claude and appreciated their patience while the team investigated.
Still, a green status page did not mean every user was back online. The article referenced a post from a user named elstar on X and said some people continued to face authentication residue and front-end issues. In those cases, the settings, usage and skills pages appeared as blank white screens even while the official system status showed normal operation. Complaints such as 「Claude down again」 continued into late night, according to the report.
August incident records were densely packed
The article cited third-party monitoring service StatusGator as saying Claude had accumulated 184 incident records from January 2026 to the present, even though the calendar had only reached August.
It also laid out a dense August incident timeline:
- Two incidents on Aug. 20.
- One incident each day from Aug. 19 through Aug. 16.
- About four hours of downtime for Fable 5 on Aug. 15.
- Three incidents in a single day on Aug. 14.
- One incident each on Aug. 13 and Aug. 12.
The article said the most severe outage in that stretch came on Aug. 5, when the disruption lasted 7.5 hours. Users then saw the message 「Due to unexpected capacity constraints.」
Users tied repeated errors to compute limits
The phrase 「capacity constraints」 appeared more than once in error messages during the month, according to the article. On X, the phrase users repeated was 「ran out of compute.」
One developer, as quoted in the report, said Opus 4.6 used to be strikingly capable, but that 「they seem to have completely run out of compute, and then weakened everything.」 Another user, Eason, said Claude had shown what he viewed as an unprecedented decline in capability, adding that Opus 5 had become unusable for him and that he had gone back to Opus 4.8.
The article also made clear that these claims were largely based on user experience rather than hard proof. It then cited one direct comparison: on the same difficult task, Opus 5 at xHigh took one hour, while a comparable model from another company finished in 15 minutes, a fourfold gap in completion time.
6,852 Claude Code sessions were reviewed in one analysis
The article also said AMD AI Director Stella Laurenzo reviewed 6,852 Claude Code sessions from her team and found a sharp drop in a tracked curve.
According to the report, the model’s reasoning depth was roughly 2,200 characters at the end of January. By early March, that figure had fallen to 560 characters, a drop of about 75%.
The same period also coincided with Claude Code gradually hiding its thinking content from users. The article said the hidden share rose from 1.5% to 100% within a single week. In the framing of the report, users were left seeing less of the model’s reasoning process at the same time they felt the model had become less capable.
Uptime figures do not capture the full cost of interrupted agent tasks
Anthropic’s own 90-day uptime figures, as cited in the article, were 99.33% for claude.ai, 99.43% for the API and 99.35% for Claude Code. The report argued that those numbers may look high at first glance, but that enterprise-grade services are often measured against a 99.9% line.
It also drew a distinction between traditional web interruptions and agent-style workloads. In a conventional internet service, an outage may simply mean a page fails to load and a user refreshes a few seconds later. For an agent handling a long-running task, though, an interruption can erase not only the minutes of downtime but also the time, context and compute already spent before the failure. The article argued that a status page cannot fully record that loss.
As of the time described in the article, Anthropic still had not explained what exactly happened during the three-hour disruption on Aug. 24. August, meanwhile, had already logged 13 incident days, with seven days still left in the month.