Re:Is Opus 5.5 nerfed? Thinking patterns changed on November 3
by xnorswap on Monday November 09, 2026 @06:47AM
(#50424290)
At this point, my own experience has been so diverged with the online chatter that I need to consider that either:
1. I'm in a A/B with a ruthlessly effective model
2. I've tuned my memory files in a way the model likes
3. I've been lucky
4. I'm just "better at claude" than most
It's impossible to say, and that is clearly part of the problem. More visibility of this stuff might not even help us, given it is human nature to ascribe patterns to randomness where they don't exist.
Sounds like analysis fully driven by a model, which can be fine, if you know what you are getting into and can verify the output. Just last week had my dad ask Claude Opus 5.5 on High about watermarking after I informed him. The model told him that watermarking text was not implemented in Anthropic models, whereupon he showed me and I then asked him to just ask the model to check again, ideally online. These models do still "hallucinate", they still have major gaps in their output not easily caught if you don't know the topic and interestingly, anything connected to LLMs, model providers, etc. tends to be where they do worst.
In this case, looks like the LLM did not consider that these are (and have been for months now) just summaries and the actual reasoning can not be analyzed for encryption reasons.
If in doubt, please do some research independent of the model output. LLMs not having any internal "understanding" that'd give them anything akin to an internalized "feel" for what they "know" accurately enough is one of their major flaws.