TL;DR — Anthropic Postmortem of Three Recent Issues

In Aug–Sep 2025, Claude users saw degraded output quality due to infrastructure bugs, not intentional changes.

The Three Issues 1. *Context window routing error* - Short-context requests sometimes routed to long-context servers.

   - Started small, worsened after load-balancing changes.

2. *Output corruption* - TPU misconfigurations led to weird outputs (wrong language, syntax errors).

   - Runtime optimizations wrongly boosted improbable tokens.

3. *Approximate top-k miscompilation* - A compiler bug in TPU/XLA stack corrupted token probability selection.

   - Occasionally dropped the true top token.

Why It Was Hard to Detect - Bugs were subtle, intermittent, and platform-dependent.

- Benchmarks missed these degradations.

- Privacy/safety rules limited access to real user data for debugging.

Fixes and Next Steps - More sensitive, continuous evals on production.

- Better tools to debug user feedback safely.

- Stronger validation of routing, output correctness, and token-selection.

▲ sebastiennight 4 days ago | parent [-]

> Privacy/safety rules limited access to real user data for debugging.

Do their ToS really limit access to user data (prompt/response)? I don't remember seeing anything to that effect in their terms.

	▲	mcintyre1994 4 days ago \| parent \| next [-]
		I’d imagine they have a lot of internal controls, even if ultimately someone at the company can read the data within their terms. It makes sense that the teams debugging stuff wouldn’t have this access immediately.
	▲	favorited 4 days ago \| parent \| prev [-]
		I know that when you submit a thumbs up/down rating for a response, you need to opt-in to the whole chat conversation being shared with Anthropic.