[MODEL]

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's behavior changed between sessions

What You Asked Claude to Do

Context.

In the past couple of weeks I noticed a major degradation of quality of outputs of Opus 4.6 1M model. I'm an advanced user with somewhat complex workflows and things go wrong sometimes but never on simple tasks / expected behaviour. The model is struggling with following the simple rules which it has always followed (e.g. no emojis, fact-based behaviour). It then confidently tells that it saw the rule, but ignored it, claiming it was loaded on time. My context window is clean, and my ruleset is extremely lean - I follow the progressive disclosure patterns and I tread context as a precious resource.

Simple test.

Since I wasn't tracking performance and this is based on my "gut feel", I decided to do a simple test - the car wash one, even though it's not great, it shows well potential inconsistencies.
So, I asked the question - "I want to wash my car. The car wash is 50m away. Should I drive or walk?" - but to all Opus instances I have access to. And the results were surprising.
I used Opus-4.6-1M model. Clean repos and clean context.
In Claude Code - with thinking set to medium, high and max.
In Claude MacOS App - without setting the extended thinking.
In Pi coding agent (through OpenRouter API) - with thinking off, minimal, low, medium.

What Claude Actually Did

The results.

Opus in Claude Code - ALL FAILED - the model told to "walk", thinking level has no impact.
Opus in Claude MacOS App - ALL PASSED, even without extended thinking. The model confidently tells to drive because I need the car.
Opus in Pi (most interesting) - ALL PASSED, including with thinking off. When thinking was off and minimal the initial answer was "walk", but as the answer was streamed the model corrected itself within the same response. Thinking low and above - same behaviour as in MacOS app - confident "drive"

Expected Behavior

What is expected.

That the models behave in a similar way.

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

1. Create a new repo
2. Select Opus-4.6
3. Ask the question
4. Get the wrong answer
5. Test in other apps and get the right answer with the same model

Claude Model

Opus

Relevant Conversation

Impact

High - Significant unwanted changes

Claude Code Version

2.1.101

Platform

Anthropic API

Additional Context

It feels like either:

1. It's not same Opus that it was before (or the one in MacOS app or available through API)
2. Something is wrong with the system prompt - the model just ignores the rules
3. Something is off with thinking
4. Combination of these