On September 3, 2026, at around 9:41 AM ET, Anthropic reported an outage that produced elevated errors for requests to multiple Claude models.
Timeline: detection, identification, and ongoing remediation
Anthropic’s status updates show the incident began at about 9:41 AM ET on September 3, 2026, when the company said it was investigating elevated errors on requests to Claude Mythos 5.1, Claude Fable 5.1, and Claude Opus 5. At the same timestamp, Anthropic reported it had identified the cause of the elevated errors and was working on a fix. A short time later the company published a more complete list of affected models. By 10:49 AM ET Anthropic said it was continuing to work on a fix, and the outage remained ongoing at the time of writing.
Which Claude models are affected
Anthropic confirmed that multiple iterations of its Claude family were impacted. The models the company listed as affected are:
- Mythos 5.1 and Mythos 5
- Fable 5.1 and Fable 5
- Opus 5, Opus 4.8, and Opus 4.6
The company’s updates moved from a narrower initial set (Mythos/Fable 5.1 and Opus 5) to a broader roster a few minutes after detection, indicating the problem affected multiple model versions across the Claude lineup.
Audit-ready is a season. It shouldn't be.
Evidence in spreadsheets, controls drifting between audits, frameworks multiplying on flat headcount. Nubivance runs continuous compliance on Rapid7 Cyber GRC - SOC 2, HIPAA, ISO 27001, PCI, CMMC.
End the scrambleUser impact: failing requests and returned errors
For people and systems using the affected Claude models, Anthropic’s advisory says the outage can cause requests to fail to complete or to return errors. The status page framed the problem in operational terms: elevated error rates on request paths to those models, producing partial or failed responses for users attempting to reach Claude Mythos, Fable, or Opus instances listed above.
Security context: prevention scores and the Blue Report 2026
The source material also includes a security observation: “Overall prevention scores can hide what happens after initial access. Once attackers are using valid credentials, prevention drops sharply.” That point appears alongside a reference to the Blue Report 2026, which the notice describes as a measurement of defenses “technique by technique” across 338 million simulations run in customer production environments. The Blue Report 2026 is presented as a diagnostic that differentiates prevention at initial access from resilience after an adversary has legitimate credentials.
What this means for security teams, affected enterprises, and end users
- Security teams: The notice’s inclusion of the prevention-score observation underscores the distinction between blocking initial incidents and managing impact once access exists. Teams monitoring ML platforms should note that elevated service errors can complicate incident response and forensic collection when model endpoints are unstable.
- Affected enterprises and procurement leaders: Organizations that rely on the listed Claude versions may experience partial service loss or degraded automation that depends on completed model responses. The staggered disclosure of affected versions suggests operators should verify which specific model versions their systems call and monitor vendor status pages closely during outages.
- End users: Individuals using Claude-powered applications may see requests fail or receive error messages when interacting with the named Mythos, Fable, or Opus models; users should expect intermittent failures until Anthropic completes its remediation work.
The record in Anthropic’s status updates is straightforward: an outage was detected at roughly 9:41 AM ET on September 3, 2026; the company identified a cause and began work on a fix shortly thereafter; a broader set of models was confirmed affected minutes later; and as of 10:49 AM ET the company was still working on resolution and service remained degraded. The status page, the list of impacted model versions, and the posted times are the elements that define both the scope of impact and the vendor’s response timeline.
