Anthropic Explains Recent Claude Issues
By Anthropic has shared a report on three infrastructure bugs that hurt Claude’s response quality during August and early September. The problems have been fixed, and the team outlined what went wrong and what they are changing going forward.
What Happened
-
Routing mix-up: Some Sonnet 4 requests were sent to servers prepared for a new 1M context window. This caused weaker replies for part of the user base until the routing was corrected in early September.
-
Output errors: A misconfiguration on TPU servers led to occasional corrupted text, such as random characters or broken code appearing in answers. The issue was rolled back within a few days.
-
Compiler bug: A flaw in Claude’s TPU compiler affected token selection for certain models. The change was reverted and replaced with a more precise approach.
Why It Was Hard to Catch
Because the three issues overlapped and showed up differently across platforms, reports looked inconsistent and took longer to trace back. Anthropic noted that model quality is never reduced based on demand or load and that these incidents were caused only by infrastructure faults.
What’s Changing
The team is rolling out more sensitive evaluations, continuous monitoring on live systems, and faster debugging tools that still protect user privacy. They also pointed out that direct user feedback was key in spotting the problems.
Takeaway
Claude’s quality should now be back to normal. The report highlights the scale and complexity of running large models across different hardware, and how vital it is to keep users closely involved in monitoring.