Anthropic Postmortem: Three Infrastructure Bugs Degraded Claude's Responses With No Change to the Model (August-September 2025)

Summary

In a postmortem published on September 17, 2025, Anthropic said that three separate infrastructure bugs intermittently degraded the quality of Claude's responses between August and early September 2025, and that the problems were due to infrastructure bugs alone, not to any reduction in model quality. By Anthropic's account: from August 5 some Sonnet 4 requests were misrouted to servers configured for an upcoming 1M-token context window, rising after a routine load-balancing change on August 29 to 16% of Sonnet 4 requests at the worst hour on August 31, and about 30% of Claude Code users had at least one misrouted message; an August 25 misconfiguration on TPU servers corrupted output, for example inserting Thai or Chinese characters into English answers or syntax errors into code; and an August 25 change to token selection triggered a latent compiler bug affecting Claude Haiku 3.5 and possibly some Sonnet 4 and Opus 3 requests. Anthropic said users began reporting degraded responses in early August, that the reports were hard to tell from normal variation, and that it opened an investigation in late August. It said its evaluations did not capture the degradation, that it did not at first connect the August 29 spike in reports to the load-balancing change, and that privacy controls limited engineers' access to affected conversations. It rolled back or fixed each change between September 2 and September 18 and said it would run quality evaluations continuously on production systems. These facts are as Anthropic reported them.

Consequence

• Anthropic said it rolled back the output-corruption change on September 2 and added detection tests for unexpected character outputs to its deployment process
• Anthropic said it deployed the routing fix from September 4, completing the rollout by September 18
• Anthropic said it will run quality evaluations continuously on production systems

Who caught it A client or customerFound outside the organization

Key dates

1. August 2025Event
2. October 21, 2026Checked against sources

Developments are added when the record is reviewed. Next review: October 21, 2027.

Evidence

Each passage is quoted from its source. The quoted passages are re-checked against the live pages every month.

1. Between August and early September, three infrastructure bugs intermittently degraded Claude's response quality. We've now resolved these issues and want to explain what happened.

   Supports: what and when: three infrastructure bugs degraded Claude's response quality, August to early September 2025

   To state it plainly: We never reduce model quality due to demand, time of day, or server load. The problems our users reported were due to infrastructure bugs alone.

   Supports: Anthropic's statement that the problems were due to infrastructure bugs alone, not a reduction in model quality

   In early August, a number of users began reporting degraded responses from Claude. These initial reports were difficult to distinguish from normal variation in user feedback.

   Supports: how caught: user reports from early August, hard to distinguish from normal variation

   By late August, the increasing frequency and persistence of these reports prompted us to open an investigation that led us to uncover three separate infrastructure bugs.

   Supports: the reports prompted an investigation in late August that found the bugs

   On August 5, some Sonnet 4 requests were misrouted to servers configured for the upcoming 1M token context window

   Supports: bug 1: context-window routing error from August 5

   At the worst impacted hour on August 31, 16% of Sonnet 4 requests were affected.

   Supports: scale: 16% of Sonnet 4 requests at the worst hour

   Approximately 30% of Claude Code users who made requests during this period had at least one message routed to the wrong server type, resulting in degraded responses.

   Supports: scale: about 30% of Claude Code users had at least one misrouted message

   On August 25, we deployed a misconfiguration to the Claude API TPU servers that caused an error during token generation.

   Supports: bug 2: output corruption from an August 25 misconfiguration

   for example producing Thai or Chinese characters in response to English prompts, or producing obvious syntax errors in code.

   Supports: symptoms of the output corruption

   On August 25, we deployed code to improve how Claude selects tokens during text generation. This change inadvertently triggered a latent bug in the XLA:TPU

   Supports: bug 3: token-selection change triggered a latent XLA:TPU compiler bug

   Resolution: We identified the issue and rolled back the change on September 2. We've added detection tests for unexpected character outputs to our deployment process.

   Supports: consequence: output-corruption change rolled back September 2; detection tests added

   We deployed the fix on September 4. Rollout to our first-party platform and Google Cloud's Vertex AI was completed by September 16, and to AWS Bedrock by September 18.

   Supports: consequence: routing fix deployed from September 4, completed by September 18

   The evaluations we ran simply didn't capture the degradation users were reporting, in part because Claude often recovers well from isolated mistakes.

   Supports: Anthropic's account of why its evaluations missed the degradation

   When negative reports spiked on August 29, we didn't immediately make the connection to an otherwise standard load balancing change.

   Supports: the August 29 spike was not immediately linked to the load-balancing change

   Our internal privacy and security controls limit how and when engineers can access user interactions with Claude, in particular when those interactions are not reported to us as feedback.

   Supports: privacy controls limited engineers' access to the affected interactions

   Although we run regular evaluations on our systems, we will run them continuously on true production systems to catch issues such as the context window load balancing error.

   Supports: commitment: continuous quality evaluations on production systems

   Published Sep 17, 2025

   Supports: postmortem publication date

   Resolution: We first observed the bug affecting Haiku 3.5 and rolled it back on September 4. We later noticed user reports of problems with Opus 3 that were compatible with this bug, and rolled it back on September 12.

   Supports: bug 3 rollbacks: Haiku 3.5 on September 4, Opus 3 on September 12

   We also believe this could have impacted a subset of Sonnet 4 and Opus 3 on the Claude API.

   Supports: bug 3 may also have affected some Sonnet 4 and Opus 3 requests

   A postmortem of three recent issues Anthropic Engineering (September 17, 2025) Primary source