← sources

Opus 5.5 got noticeably stupid overnight.

source: reddit (reddit:ClaudeCode)published 2026-10-19 16:11ingested 2026-10-20 02:39

retained text

In just 24 hours it noticeably started:

1. Making completely false assertions.
2. Ignoring specifications, even when told exactly what to do.
3. Forgot how logic works.

Less than 24 hours ago I could tell it, "I have this idea, figure out the proper maths to make it rigorous using singular learning theory," and it just worked. Today? I can't even get it to distinguish between, "this is what I just did," and "this is a separate goal we are building towards". Nope, it just thinks, "I did something, thus the goal is achieved!"

Its abilities degenerated so much in just twenty-four hours. Anyone else experiencing this? My guess is one of (1) Anthropic pushed a bad quantization (2) off-policy distillation (3) it's becoming misaligned with more training (4) active sabotage against AI researchers.

EDIT: Figured it out. It was (4). Anthropic was silently downgrading my model to Sonnet 5 when it detected I was working on AI research. I wasn't even provided a note/warning about the downgrade, I had to ask the model who it was.

document fetched 2026-10-20 00:39 · canonical https://reddit.com/r/ClaudeCode/comments/1xa140l/opus_55_got_noticeably_stupid_overnight