Anthropic's Claude services went down hard on April 28. The outage lasted 78 minutes, from 17:34 to 18:52 UTC, and it hit everything. Claude for Government were among those hit by widespread outages. claude.ai, the API, Claude Code, Claude Cowork, and Claude for Government all had issues. Users couldn't log in. API requests failed with authentication errors. Anthropic confirmed resolution by 19:15 UTC.
One outage isn't the real story. The pattern is. A single Hacker News commenter, claiming to represent an enterprise customer spending over $200,000 monthly, reports just "one 9 of uptime" over the past 90 days. That's roughly 90% availability, or about 72 hours of downtime per month. If accurate, it's a huge gap from Anthropic's enterprise SLA of 99.9% uptime, which allows just 43 minutes of downtime monthly. The April 28 incident alone burned through that entire budget.
The tension is building across the AI agent space. Companies are pushing tools like Claude Code for production work, betting engineering workflows on services that can't consistently stay up. The SLAs match traditional cloud contracts, but the systems behind them lack the maturity of AWS or Azure. And when failures happen, compensation means service credits, not business continuity. Claude for Government being caught in the same outage raises questions about whether critical infrastructure should share failure modes with consumer products. Teams building on AI agents need to plan for outages, because they're coming.