API severe degradation since June 22 21:00 UTC — 3-18 min response times on Sonnet 4.6

Open 💬 0 comments Opened Jun 23, 2026 by joseluislucio

Summary

Experiencing severe API performance degradation since June 22 ~21:00 UTC, affecting Claude Code sessions on claude-sonnet-4-6. Response times that normally take 5-30s have been ranging from 1.5 to 18.6 minutes per turn.

This predates the currently-open incident #2stpj60m2fr9 ("Elevated errors for Claude Opus 4.8", started June 23 06:28 UTC) and affects Sonnet 4.6, not just Opus 4.8.

Environment

  • Claude Code version: 2.1.186 (installed June 22 ~22:48 UTC)
  • Model: claude-sonnet-4-6
  • Platform: macOS (Apple Silicon)
  • API latency to api.anthropic.com: ~70ms (network is fine)

Measured API response times (user→assistant gaps from session transcripts)

Session 1 (b00a14c6, June 22 21:00 – June 23 07:31 UTC, Boxto project)

| Time (UTC) | Response time |
|---|---|
| 21:38→21:42 | 213s |
| 21:51→21:57 | 304s |
| 22:03→22:11 | 502s (8.4 min) |
| 22:12→22:20 | 487s |
| 22:25→22:32 | 460s |
| 22:32→22:42 | 564s (9.4 min) |
| 23:06→23:12 | 367s |
| 23:24→23:31 | 435s |
| 06:19→06:37 | 1115s (18.6 min) |
| 07:19→07:26 | 418s |

Average over 59 measured gaps: 192 seconds. Maximum: 1115 seconds.

Session 2 (0b812c45, June 23 10:37–11:04 UTC, same project, new session)

After the reported fix at 07:47 UTC, degradation continues:

| Time (UTC) | Response time |
|---|---|
| 10:37→10:40 | 180s |
| 10:43→10:45 | 159s |
| 10:46→10:48 | 137s |
| 10:51→10:53 | 111s |

The fix applied at 07:47 UTC has not resolved the issue for Sonnet 4.6.

Impact

  • Tasks that should take minutes take 1-2 hours
  • Sessions accumulate enormous CPU time waiting on API calls (one session: 40+ minutes of CPU over 10 hours)
  • The status page incident (#2stpj60m2fr9) appears to underreport: started at 06:28 UTC June 23, but degradation was present at least since 21:00 UTC June 22 (~9.5 hours earlier)

What I checked

  • Network latency to api.anthropic.com: 68ms ✓ (not a network issue)
  • Local system resources: adequate ✓
  • Hooks overhead: <1s ✓
  • Model: claude-sonnet-4-6 (not Opus 4.8 which the incident references)

Please check whether the degradation scope and start time are accurately reflected in the active incident.

View original on GitHub ↗