I'm a daily user of Claude Code and love it. However, the quality of responses (using Opus 4.1) seems to have noticeably degraded over the last week.
Unfortunately, I don't have a good way to objectively measure quality aside from "it just feels like it's gotten worse" so thought I'd ask HN if others have noticed a change.
It has gotten worse. As a daily user myself I can tell you that the performance degradation, at least for me, began somewhere after 20th August, with scrappy results in the web interface and in the cli (not following instructions, forgetting instructions, to-do list not working, not following claude.md orders, code artifacts being half-done or not rendering at all and so on).
Yes. It was noticeable to me through daily usage. So much so that I checked anthropic website for any issues they might have posted about. Opus limits also seemed to have been tightened.
I also feel that the responses and actions have gotten slower around the same time. It seems like there is a larger delay between messages and tool executions.
I have actually, I'm glad you made this. Was thinking offhandedly it might be a result of rushed changes due to their lawsuit with NYT. But I've noticed it get worse on my model (claude-sonnet-4-20250514) over the last month, meaning it could be an issue with how the CLI prompts the model? My third conspiracy would be that they somewhat quantized the models to save cost. I saw that they had an outage after trying to use trainium chips for inference, perhaps they are pitching wild to try and keep up with their margins
