Claude's Aug 24 Outage: Elevated Errors Hit Opus 5, Fable 5, Mythos 5 for 2h46m
Claude's Aug 24 Outage: Elevated Errors Hit Opus 5, Fable 5, Mythos 5 for 2h46m
Did your Claude Code session suddenly start retrying everything around 12:50 PM Beijing time on August 24?
It wasn't just you. For nearly three hours, Claude requests across the board were failing — claude.ai, the API, Claude Code, and Claude Cowork all took a hit. Anthropic's own status page called it "Elevated errors for multiple models."
The window ran from 04:50 UTC to 07:36 UTC — 2 hours 46 minutes — and it wasn't an Opus-5-only problem. Mythos 5, Fable 5, and Opus 4.8 got hit too. If you run Claude in a production pipeline, consider this another nudge: a single-vendor dependency is a single point of failure.
The Timeline, Straight From the Status Page
Anthropic's status page (status.claude.com) logged the whole incident with timestamps. Here's the sequence (all times UTC):
| Time (UTC) | Stage | Verbatim from status page |
|---|---|---|
| 04:50 | Start | Elevated errors begin on Claude requests |
| 05:06 | Investigating | "We are investigating elevated errors on requests to Claude Mythos 5, Claude Fable 5, Claude Opus 5, and Claude Opus 4.8." |
| 05:27 | Identified | "We have identified the cause of elevated errors on requests to Claude Mythos 5, Claude Fable 5, Claude Opus 5, and other Claude models" |
| 06:42 | Update | Still working on "issues causing elevated requests on multiple models" |
| 07:36 | End | Errors stop (00:36am PT) |
| 07:47 | Update | "errors stabilize on Opus 5 and Fable 5, and are working to fully resolve success rates on all affected models" |
| 08:30 | Resolved | "This issue has been resolved." |
The incident summary, verbatim:
"From 9:50pm PT / 04:50 UTC through 00:36am PT / 07:36 UTC, users saw elevated errors on requests to Claude models, including Claude Opus 5 and Fable 5."
Scope: Models and Products Affected
Few incidents have an impact list this long. From the status page:
- Models: Claude Mythos 5, Claude Fable 5, Claude Opus 5, Claude Opus 4.8
- Products: claude.ai, Claude API (api.anthropic.com), Claude Code, Claude Cowork
That's the whole surface: web chat, API calls, terminal agents, and the enterprise collaboration tool. If you were running a batch job in Claude Code, you were probably staring at a wall of 529/503s and timeouts.
The Detail That Matters: Diagnosis Was Fast, Recovery Wasn't
Here's the interesting bit: Anthropic started investigating at 05:06 and had identified the cause by 05:27 — 21 minutes. Yet the Resolved stamp didn't land until 08:30, three hours later. Even the 07:47 update still said they were "working to fully resolve success rates on all affected models."
To me this is a textbook observability case: fast diagnosis ≠ fast recovery. Design your alert thresholds, retry logic, and circuit breakers for outages lasting hours, not minutes.
Three Things to Do Now
- Don't bet your pipeline on one model vendor. Opus 5, Fable 5, and Mythos 5 went down together — this wasn't a single-model blip. Have a fallback path for critical flows (a secondary model or a local one).
- Cap your retries with backoff. Claude Code has built-in retries, but if you write your own pipeline, use exponential backoff plus circuit breaking. Don't burn through your quota inside the failure window.
- Subscribe to status-page incident alerts. status.claude.com supports email/SMS notifications, which arrive faster than any third-party monitor.
FAQ
Q: How long did the Claude outage last? A: Per the official status page, elevated errors ran from 04:50 UTC to 07:36 UTC on Aug 24 — 2 hours 46 minutes — and the incident was marked Resolved at 08:30 UTC.
Q: Which Claude models were affected? A: Claude Mythos 5, Claude Fable 5, Claude Opus 5, and Claude Opus 4.8, across claude.ai, the Claude API, Claude Code, and Claude Cowork.
Q: What caused the outage? A: The status page only states the cause was identified ("We have identified the cause"); the root cause wasn't made public. Watch for an official postmortem.
✨ 本文由 DeepSeek 生成初稿，Claude 审核润色。
Sources
- Claude Status — Elevated errors for multiple models
- Anthropic Newsroom — News