›_ Claude Code Issues
Claude Code Issues / #47512

Claude Code availability depends on claude.ai auth — API backend outages masked by frontend dependency

Resolved 💬 3 comments Opened Apr 13, 2026 by cnighswonger Closed May 25, 2026

Summary

Claude Code authenticates subscription users (Pro, Max) through claude.ai's OAuth flow. When claude.ai experiences an outage, Claude Code becomes unavailable even if the underlying /v1/messages API backend is healthy and serving requests normally.

Observed today (2026-04-13): claude.ai returned 500 errors. The model-serving API was operational (API-key authenticated requests would have succeeded), but Claude Code subscription sessions received 500s because the auth path routes through claude.ai infrastructure. The outage duration exceeded the cache TTL on at least one active agent session, triggering a full cold-cache rebuild on recovery.

Impact

  • Subscription users (Pro, Max 5x, Max 20x): completely blocked during claude.ai frontend outages, even for existing authenticated sessions
  • API-key users: unaffected (auth is direct to the API)
  • Cache cost: if the outage outlasts the cache TTL (5m or 1h depending on tier), users pay a full cache rebuild on recovery — potentially 100K–900K tokens of cache creation for long-context sessions. We observed this firsthand: at least one active agent session suffered a cold-cache rebuild because the outage outlasted its TTL window.

Expected behavior

Claude Code sessions that are already authenticated should be able to continue making API calls during claude.ai frontend outages, since the model-serving backend is a separate service.

Suggested mitigations

  • Token caching / refresh resilience: If the OAuth token is already issued and valid, Claude Code should be able to continue using it without re-validating against claude.ai on every call
  • Graceful degradation: If auth refresh fails due to a claude.ai outage, continue operating on the existing token until it actually expires rather than failing immediately
  • Status page separation: Distinguish claude.ai frontend status from API/Claude Code status on status.anthropic.com so users know which service is actually down

Related

  • #38335, #42052, #45756 — quota drain complaints; outage-triggered cache rebuilds compound the cost problem documented in those issues
  • The claude-code-cache-fix /coffee keepalive skill mitigates the cache-rebuild cost on recovery but can't mitigate the outage itself

View original on GitHub ↗

This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗