very undependable; unpredictable; losing confidence; new models are deteriorating

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

I am using a premium license. Not sure whether i am getting money's worth. Just losing peace of mind.

There was a clear plan, which was generated after mulitiple rounds of plan generation in plan mode. This was to create a EVALS for an accelerator that I am developing

The moment it started executing, there were several issues and I have taken screen shots of all its behaviour. I am attaching them. Are we expecting a lot from these models? They don't behave like a matured model. I dont mind sharing the plan either.

EVALS-PLAN.md

<img width="616" height="227" alt="Image" src="https://github.com/user-attachments/assets/4f2f9430-a63f-46cb-87b3-943c911ba036" />
<img width="575" height="737" alt="Image" src="https://github.com/user-attachments/assets/94708467-faa8-4a76-86c7-7d1e8e8321bb" />
<img width="555" height="646" alt="Image" src="https://github.com/user-attachments/assets/0c958c7e-e418-4a8a-a604-adfa278b5461" />
<img width="566" height="618" alt="Image" src="https://github.com/user-attachments/assets/abc915ce-22a8-446e-b967-dd10822844b8" />
<img width="592" height="691" alt="Image" src="https://github.com/user-attachments/assets/49ad48f0-58c1-4933-a500-ca1f857e3bec" />

What Claude Actually Did

1. The plan it created, it could not execute
2. handed it back to me
3. when pushed it started executing
4. reviews threw a lot of issues (normal behaviour)
5. asked it to fix - again handed it back to me saying Manual escalation
6. when questioned, went an executed more than what it should.
7. i had asked for taking approval from after every stage is completed. went berserk and completed all the stages
8. But missed fixing 3 issues in the stage 1 - originally what I had asked it to do

very erratic and unpredictable

Expected Behavior

follow the plan
follow the instructions
do what it is supposed to do than pushing back

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Haven't tried to reproduce

Steps to Reproduce

_No response_

Claude Model

Opus

Relevant Conversation

Impact

High - Significant unwanted changes

Claude Code Version

2.1.238 VS code extension

Platform

Other

Additional Context

_No response_