Claude went down for 85 minutes yesterday. Government customers never noticed.

Anthropic's Claude platform went dark on Tuesday.

8,000+ Downdetector reports in the first hour.

Every commercial platform hit simultaneously: claude.ai, the API, Claude Code, Console, Cowork.

But Claude for Government? Untouched.

It runs on separate FedRAMP infrastructure inside Palantir's federal cloud. 99.93% uptime over 90 days.

The commercial version you're building on? 99.12% uptime.

That's 19 hours of downtime per quarter. Enterprise SLAs demand 99.9%. You're getting 10x less reliability than your contracts require.

This is the third major Claude outage in June.

The June 2 failure traced to a sub-agent architecture bug in Claude Code that spawned infinite loops and crashed the entire platform. CI/CD pipelines that routed code review through Claude didn't pause — they failed the run, requiring manual cleanup from scratch.

Thoughtworks framed it directly: "AI tools should amplify engineers' capabilities. It shouldn't act as a structural crutch."

Your move:

Audit every production system that calls Claude. If your CI/CD pipeline has no fallback when the API returns HTTP 500 or 529, you're one outage away from a manual restart.

Treat Claude's direct API and AWS Bedrock as separate failure domains. They run on different infrastructure. One fails, the other survives.

Add circuit breakers to every Claude-integrated workflow. Degrade to queued review. Don't fail the run.

Single-vendor AI dependency is now a documented single point of failure. Plan for the model being down before it happens.