Claude's rough morning: Anthropic logs elevated errors across five models, twice before dawn

• Who: Anthropic, on its Claude status page.
• What: Elevated errors on requests to five Claude models — Mythos 5.1, Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5.
• When: First incident posted at 06:14 UTC (2:14 a.m. Eastern) on October 25, 2026; impact ran 06:05–06:35 UTC.
• Where it showed up: claude.ai, the Claude API, Claude Code and Claude Cowork were marked as partial outages; Claude Console and Claude for Government stayed operational.
• Second incident: Elevated errors for Claude Fable 5.1, impact 08:08–08:29 UTC (21 minutes), resolved 4:49 a.m. Eastern.
• Context: The status page also lists three resolved incidents on October 24, 2026, and 90-day uptime of 99.58% for claude.ai and 99.64% for the Claude API.

Anthropic's Claude had a rough start to Sunday. In the early hours of October 25, 2026, the company's status page logged elevated errors across five Claude models at once, marking the main ways people reach Claude — the website, the API, the coding tool and the Cowork app — as partial outages. The impact window was short, about half an hour, and a fix was deployed within 21 minutes of the first post. Then, a few hours later, a second incident hit one model again. For anyone running Claude inside a product or a workflow, the morning is a reminder that even a 30-minute wobble lands hardest on the systems that assume the model is always there.

The first incident: five models, four services

Anthropic posted the incident, titled “Elevated errors for multiple models,” at 06:14 UTC — 2:14 a.m. Eastern — listing elevated errors on requests to Claude Mythos 5.1, Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5. The affected services were claude.ai, the Claude API, Claude Code and Claude Cowork, all marked as partial outages. Claude Console and Claude for Government stayed operational.

The timeline moved fast. At 06:27 UTC Anthropic said it had identified the cause of the errors on Sonnet 5.5 and was working on a fix; by 06:35 UTC a fix was deployed and the incident was in monitoring. At 07:53 UTC the incident was marked resolved, with the stated impact window of 06:05 to 06:35 UTC on October 25 — 30 minutes from first impact to fix, even though the formal resolution came later, once recovery was confirmed.

A second incident later the same morning

Claude was not done for the day. A second incident, titled for elevated errors on Claude Fable 5.1, was posted at 4:17 a.m. Eastern, with the cause identified in the same update. A fix was deployed by 4:30 a.m. Eastern and the incident was marked resolved at 4:49 a.m. Eastern, with impact stated as 08:08 to 08:29 UTC — 21 minutes. The confirmed problem in that incident was limited to Fable 5.1, a narrower scope than the five-model incident hours earlier.

Two things stand out in the status updates. First, Anthropic said it had identified the cause in each case but did not publish the underlying cause in the updates quoted here — so this article does not speculate about one. Second, the pattern of the morning was fast detection and fast fixes, not a long outage: both impact windows were under 35 minutes.

A pattern, not a one-off

The October 25 incidents did not arrive out of nowhere. The status page's recent history lists three resolved incidents on October 24, 2026, including one for Fable 5 and Fable 5.1 and one for Opus 5.5, plus a Connectors incident on October 22 and degraded performance for Fable 5 on October 21. Taken together, the last few days read less like a single bad morning and more like a run of small, fast-moving reliability wobbles.

The status page's 90-day uptime figures put the wobbles in proportion: 99.58% for claude.ai, 99.64% for the Claude API, 99.58% for Claude Code and 99.57% for Cowork, against 99.84% for Claude Console and 100% for Claude for Government. In other words, the service is up the overwhelming majority of the time — but the failures that do happen tend to cluster, and they tend to hit several models at once.

What it means if you build on Claude

For individual users, a 30-minute window of elevated errors is an annoyance: retries, a reload, a coffee. For developers and businesses, it is a design question. Agentic workflows, coding assistants and customer-facing tools that call Claude without retries, fallbacks or a status check will fail silently or noisily in exactly these windows — and the failure will look, at first, like a bug in your own code. The practical defences are boring and cheap: retry with backoff, surface the provider's status before blaming your prompt, and keep a degraded mode that does not require a live model call.

The other lesson is about where to look. Anthropic's own status page — status.claude.com — carried the incident, the timeline and the resolution, and secondary coverage from Unite.AI and Technobezz tracked the same updates. When a frontier model starts erroring, the status page is the fastest way to tell a provider-side incident from a problem on your side. The October 25 morning gave that advice two chances to prove itself before breakfast.