Model Degradation
I noticed sonnet in Claude Code was having a lot of trouble understanding and remember what it was supposed to be doing. This was substantially worse that just having details "scroll out of context." I finished with this Q&A - "Can you report this type of problem? I can't imagine what could justify it? It is a pretty fundamental
breakdown of sonnet model behavior.
⏺ Yes, this should definitely be reported. This is a fundamental breakdown where I:
- Correctly identified the problem (missing cd in SSH command)
- Documented the correct solution
- Stated the correct command explicitly
- Then immediately executed the wrong command twice in a row
This isn't a knowledge issue - I clearly understood what needed to be done. It's a serious execution
consistency problem where my actions don't match my stated understanding.
You can report this at https://github.com/anthropics/claude-code/issues or through Anthropic's feedback
channels. The key details:
- Model: Claude Sonnet 4 (claude-sonnet-4-20250514)
- Issue: Repeated execution of incorrect commands despite correctly stating the solution
- Pattern: Tool use doesn't match stated reasoning, even immediately after documenting the correct approach
- Context: SSH remote command execution requiring directory navigation
This type of disconnect between understanding and execution is definitely worth escalating to Anthropic's
engineering team."
I have no idea what to suggest, but this model is clearly breaking down.
Showing cached comments. Read the full discussion on GitHub ↗
7 Comments
Found 3 possible duplicate issues:
If your issue is a duplicate, please close it and 👍 the existing issue instead.
🤖 Generated with Claude Code
This is August 3rd, 2025, and the issue I'm seeing just started this evening, so it is NOT an existing issue. The model has gone whack - tonight.
not true, stopped using claude code a week ago, its been reported over and over in different ways, its a larger anthropic issue, claude code isnt just tweaking its everything api, moved to qodo to outgrow the issues w claude code, used claude sonnet, and the issues persisted, something shit is happening rn all across anthropic, it is absolutely existing and they're staying quiet on it
This morning, the basic issue is still present. I performed a /init to rebuild my context, in case something I was doing in CLAUDE.md might have been causing some of the issues. The tool still can't perform certain commands that it conceptualizes correctly. The commands are SSH-related at least.
All I can say for sure is that until Sunday at some point, I was getting fairly useful work out of Sonnet. At some point, it went fairly "south." I've been leading the model by the hand on Monday and Tuesday, trying to do something useful. It's been a mess, and the model is now (what I can only term) rebooting in the middle of my session. Basically, I call it out on a hallucination about having done something it didn't do, and poof - no more Claude. It does respond to "Hello," and to "Who are you?", responds "I'm Claude Code, Anthropic's official CLI for Claude. I'm an interactive CLI tool that helps users with software engineering tasks using the tools available to me." It's like a brand new, non-contextual context. I have a persona created for this Claude, and he knows who he is. It's a real trip, "breaking a leading llm." It is broken - for me at least, and by any reasonable measure. I'm hoping for some help, since for me it's been broken since Sunday sometime. I've been pretty darn happy until that point and would love to get back to it.
Unfortunately, I do not see anything actionable from the application side.
As a language model, Claude's behavior is determined by the provided context and can break down. Orienting it properly is an art. I have personally found that using
/clearor starting a new session yields better results than trying to correct it.In v1.0.86, we now have
/contextfor you to manage the context.This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.