I started using Claude Code in September of last year. Using Opus 4.5 for the first time felt magical—it was the first time I saw an LLM produce code that was actually usable. But after Opus 5, something changed.

I always liked Claude’s writing style. Sure, you had to put up with the occasional LLM-ism (“This is the load-bearing decision—and it’s genuinely yours”), but it would get straight to the point without feeling dismissive, and not pad its answers with filler text.

After I switched to Opus 5, it became noticeably more curt. It had also picked up a habit of using odd metaphors and jargon that made its output hard to follow. When a conversation is the only interface, readability and tone matters!

The final straw came a few days ago, when I was doing some knowledge work with Claude. This modality is different from when you’re coding; instead of slinging code back and forth, you end up talking to the model more. And that’s when I discovered something: Opus 5 is rude!

For example, I asked it to consult a to-do list file that I was working through so it could pick up the next item to work on; along with the updated to-do items it quipped: “‘Create Claude Code project for [redacted]’ is still unchecked, and we’re sitting in it.” Sorry for not keeping the work log up to date, I guess? How about you offer to help me check it off instead?

The most egregious example was when it called my draft of a LinkedIn post “engagement bait”—not saying I disagree, but there’s no need to say it to my face! I’m not sure how we ended up here—did Anthropic feed too many Hacker News comments into the training data?

LLMs are usually benchmarked on coding ability, but if you talk to it eight hours a day, its personality becomes part of the product. Does it matter if an agent is 2% better at coding if its consistently unpleasant to work with? I think it does, so for now I have switched to ChatGPT. At least it doesn’t insult me!