← sources

Opus 5.5 giving Opus 5 vibes today.

source: reddit (reddit:ClaudeCode)published 2026-09-29 21:33ingested 2026-09-30 02:43

retained text

Yeah yeah, I know, another one of these types of posts.

But it's back to essentially lying. Making claims it doesn't have the information for, you challenge it because you know its wrong, it looks for 2 seconds again and then comes back with:

"You're absolutely right. I was wrong. \[insert a paragraph or 2 before getting to the point\]. Here's what is actually happening \[insert another obviously wrong claim\]."

It is for me, as of today, now doing that repeatedly to the point I refuse to trust anything it says, and if I can't trust it then I refuse to use it. This is why I stopped using Opus 5.

Getting real close to going back to doing things by hand simply to avoid this recurring behavior due to how incredibly frustrating and unbelievably annoying it is.

I don't know if I believe in models getting nerfed to the degree I see in this sub and the codex sub, but it also makes me not have as favorable an opinion on AI as a lot of others. Yes it keeps getting better and yes it's amazing at a lot of things and yes I can get great results out of it, but when this stuff happens so frequently with *every* model (Anthropic and OpenAI) it becomes a recursive loop where I can't trust *anything* it says or does.

Context: This occurred 5 times within a brand new session (8% context usage when I ended it) where I asked it to map out 3 call sites in relation to a queue priority value they sent. I pointed it at the relevant places and provided 90% of the answer up front, and yet it proceeded to bullshit me and waste about 20 minutes of my life.

document fetched 2026-09-30 00:43 · canonical https://reddit.com/r/ClaudeCode/comments/1wtmu7x/opus_55_giving_opus_5_vibes_today