Anthropic Claude outage Sep 3 2026: elevated errors across multiple models

Incident at a glance

Anthropic reported a multi‑model outage on Sep 3 2026 that raised error rates for Claude Mythos 5.1, Claude Fable 5.1, Opus 5, Opus 4.8, and Opus 4.6. The outage affected the web UI (claude.ai), the public API (api.anthropic.com), Claude Code, and Claude Cowork, prompting developers to fall back to older models or alternative providers.

Timeline of status updates

13:26 UTC – Investigation started – Anthropic announced they were investigating elevated errors on Mythos 5.1, Fable 5.1, and Opus 5.
13:41 UTC – Cause identified – The team said they had identified the cause of the elevated errors for the three models and were working on a fix.
13:50 UTC – Full list published – An exhaustive list of affected models was released: Mythos/Fable 5.1, Mythos/Fable 5, Opus 5, Opus 4.8, Opus 4.6.
14:49 UTC – Ongoing fix – Anthropic posted that they were continuing work on a fix.
15:25 UTC – Partial recovery – Opus 4.8 and Opus 5 remained affected; all other models had returned to baseline error rates.

Models impacted and service scope

Model	Status as of 15:25 UTC	Services impacted
Claude Mythos 5.1	Recovered (not listed in final update)	Claude.ai, API, Code, Cowork
Claude Fable 5.1	Recovered	Claude.ai, API, Code, Cowork
Opus 5	Still elevated errors	Claude.ai, API, Code, Cowork
Opus 4.8	Still elevated errors	Claude.ai, API, Code, Cowork
Opus 4.6	Recovered	Claude.ai, API, Code, Cowork

The outage was confined to Anthropic’s own infrastructure; third‑party services that depend on Claude (e.g., GitHub Copilot, X.ai’s Grok) reported similar degradation, suggesting a shared upstream failure.

Community impact and work‑arounds

Developers switched to older models – Users reported falling back to Claude Sonnet or other providers (e.g., OpenAI Codex) to continue coding tasks.
Claude Code auto‑mode broke – As noted by a commenter, Claude Code’s default auto safety mode relies on the Sonnet model. With Sonnet unavailable, auto mode could not determine safety, causing errors like:

"Error: claude‑sonnet‑5 is temporarily unavailable (overloaded), so auto mode cannot determine the safety of Edit right now."
Cross‑provider outages – Multiple commenters observed simultaneous failures on OpenAI’s Codex and X.ai’s Grok, hinting at a broader cloud‑provider incident.

"Both claude and codex down at the same time! Might actually get some quality code merged in today for a change!" – Hacker News user

"Not just claude, even grok seems to be down – https://status.x.ai/" – Hacker News user

Speculated root causes

The status page did not disclose technical details beyond “identified the cause.” Community speculation includes:

Cloudflare or upstream network issues – One commenter linked to Cloudflare’s status page, noting recent errors across multiple AI labs.
Datacenter overload – Several users pointed to “trouble in the SpaceX datacenters,” suggesting a shared infrastructure provider.
Model‑specific safety classifier failure – The Claude Code auto‑mode error indicates the Sonnet safety classifier was overloaded, which may have cascaded to other models.

No official confirmation of these hypotheses was provided.

Lessons for developers

Design for graceful degradation – Relying on a single safety classifier (Sonnet) caused Claude Code to become unusable when that model failed. Implementing fallback classifiers or disabling safety checks temporarily can keep core functionality alive.
Monitor multiple providers – Simultaneous outages across Anthropic, OpenAI, and X.ai underline the value of multi‑provider redundancy for critical workflows.
Stay aware of provider status pages – Real‑time updates from Anthropic’s status site were essential for understanding which models remained usable.

Outlook

Anthropic’s rapid identification and incremental recovery suggest a mature incident‑response process, but the recurrence of weekly errors (as reported by some users) indicates systemic reliability challenges. Developers should monitor future status updates and consider fallback strategies until the underlying cause—whether network‑level, datacenter‑level, or model‑specific—has been fully resolved.