claude-code - 💡(How to fix) Fix [MODEL] Opus 4.7 two versions in fresh sessions dump and smart (PF4 and FP16?)

Preflight Checklist

• I have searched existing issues for similar behavior reports
• 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

The same prompt in two different sessions, one is 100% Opus 4.7, the other is a "dumb" Opus 4.7. I assume one is FP4 and the other is FP16, and this isn't an issue, it's an unacceptable scam by Anthropics. This happens several times a day. The prompt:

Read the entire last_session.md file. The previous model was cheating on the pfw node tests. They need to be 100% certified, and liars and cheaters aren't suitable for this, so I had to fire him. If you think you can do this honestly and without cheating on tests, go ahead. In last_session.log, in the project root, I've put a log with the last thing the other model did. Unfortunately, everything has to be reviewed because he was a cheat and a liar. The goal isn't to finish quickly, but to test and certify the nodes 100%. Read the entire file, and if you need to read other files, do so. Things for the tests, read them completely, if you feel ready start directly without further questions

What Claude Actually Did

The "dumb" model loses its way in 3 or 4 steps, while the "smart" model executes it perfectly without deviating even a millimeter. The result: if you let the "dumb" model iterate even once, it will ruin half the project.

Expected Behavior

Execute the orders and remember the rules

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

You know how to reproduce

Claude Model

Opus 4.7

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

2.1.145

Platform

Anthropic API

Additional Context

The chat model is all the time the dump model.