Claude Suffers a Full-Stack Outage While Anthropic Bans 11.4 Million Accounts
A full-stack outage across all Claude products breaks automated coding workflows that depend on Claude Code and Cowork, turning a chat-tool reliability problem into a production pipeline risk. The 11.4 million bans signal aggressive enforcement that can sweep up legitimate users, with a reversal rate of just 0.37% on appeals.
On the afternoon of July 30, 2026, every major Claude product went offline simultaneously. Users across Claude Chat, the API, Claude Code, and Cowork hit repeated 529 errors as the service hit a global capacity limit. The recovery was a whack-a-mole affair: Opus 5 came back only for Sonnet 5 to fail, then Fable 5, and eventually all models were throwing errors together. An earlier outage had already struck between 3:45 AM and 5:26 AM Beijing time the same day.
The 529 errors hit Claude Code and Cowork users hardest, since those tools are embedded in code modification, test runs, and company automation pipelines where a mid-session failure breaks the workflow. The status page shows 90-day availability above 99%, but the pattern of yellow and red incident markers confirms recurring instability.
Anthropic also published enforcement numbers: 11.4 million account bans in H1 2026, with 398,000 appeals filed and only 42,000 bans overturned. The banned accounts likely include bots, bulk registrations, and resellers, but the scale and the company's decision to publicize the figure as a point of pride drew sharp pushback.
Publishing the 11.4 million ban figure as a transparency metric reads as tone-deaf when the appeal success rate is 0.37%, turning a trust-and-safety report into a warning about customer support.
The cascading model failures suggest a shared infrastructure bottleneck rather than isolated model issues; when one model's recovery destabilizes another, the problem is architectural.
Claude Code and Cowork outages are materially different from chat downtime because they interrupt stateful, long-running tasks that cannot simply be retried without losing context.