Claude outage: about three hours of elevated error rates
Overnight into August 24, several Claude models failed across all channels — web, apps and API. Service was back after roughly three hours.
Illustration · AI-generated (AI IN LIFE)
At a glance
- Outage on 24 August 2026: began around 4:00 UTC, resolved around 8:30 UTC — roughly three hours
- Affected: web interface, mobile and desktop apps, and the API
- Anthropic status page: "elevated errors" on multiple Claude models simultaneously
- Numerous user reports on Downdetector and X during the window
- Per media reports, not the first disruption this month
Anthropic's AI assistant Claude was only partially reachable for several hours overnight into Monday. The disruption began around 4:00 UTC and affected the web interface, the mobile and desktop apps, and the API. By around 8:30 UTC the company reported the issues as resolved — after roughly three hours of elevated error rates.
On its status page, Anthropic reported "elevated errors" on multiple models and said it was continuing to work to resolve the issues causing elevated error rates. According to the status page, several model lines were affected at once, including the current flagship models and their predecessors.
Outage trackers such as Downdetector logged a surge of user reports in parallel, with people describing error messages in chat and failed API calls. For developers and companies that have built Claude into their products, requests partially failed during that window.
According to media reports it was not the first disruption this month — the clustering comes at a time when demand for AI services, and the load on their infrastructure, is rising sharply. Other providers have seen similar incidents recently.
The episode is a reminder of a simple operating rule: anyone running AI models in production needs fallbacks — a second provider, or queue-and-retry logic for failed requests. Three hours of downtime is manageable if processes are prepared for it, and expensive if they are not.
FAQ
What exactly went down?
Several Claude models returned elevated error rates across all access paths — web, apps and API. Requests partially failed or timed out during that period.
How long did the outage last?
About three hours: from roughly 4:00 UTC until the reported resolution around 8:30 UTC on 24 August 2026.
How can companies protect themselves against such outages?
With fallback strategies: a second model provider, retry logic with queues, and monitoring of status pages so processes keep running in a controlled way during incidents.
