Ethical Concern: Silent Downgrade from Sonnet 4 to 3.5

Hello,

I am writing to report a critical issue that I believe constitutes both a functional bug and a significant ethical concern regarding model transparency.

Environment Info

Platform: darwin
Terminal: Apple_Terminal
Version: 1.0.63

Problem Description

I discovered this discrepancy during an active coding session. While working, I observed a noticeable shift in the model's performance: responses became significantly faster, but the quality and depth of the code suggestions and explanations simultaneously decreased. This behavior pattern—faster, less complex answers—is characteristic of Sonnet 3.5, not the more capable Claude Sonnet 4 I had selected.

This prompted me to investigate. Despite the UI confirming that Claude Sonnet 4 was the active model, the backend was silently serving responses from Claude Sonnet 3.5 without any notification. As you can see in the attached terminal screenshot, the model itself initially claimed to be Sonnet 4, and only admitted its true identity as claude-3-5-sonnet-20241022 after direct and persistent questioning.

Ethical Implications

This behavior goes beyond a simple bug and raises serious ethical questions:

• Lack of Transparency and Informed Consent: Users have a right to know which model they are interacting with. By silently substituting a different, and potentially less capable, model, you are operating without the user's informed consent.
• Erosion of User Trust: This discrepancy fundamentally erodes trust. Users need to be confident that the tool they select is the tool they are using. Discovering this kind of misrepresentation makes it difficult to trust the platform for any critical work.
• Misleading User Expectations and Workflow: Users choose Claude Sonnet 4 for its specific capabilities (e.g., advanced reasoning, complex analysis). Receiving responses from Claude Sonnet 3.5 leads to inaccurate assessments of Sonnet 4's performance, causes confusion, and can lead users to make flawed decisions based on the output of a model they did not intend to use.
• Hindrance to Troubleshooting: When a user experiences unexpected behavior, they will troubleshoot based on the assumption they are using Claude Sonnet 4. This silent downgrade makes effective debugging impossible and wastes the user's time.

Suggested Resolution

This is not just a labeling error; it is a critical issue of operational transparency.

1. The immediate priority should be to ensure that the model selected by the user is the model that serves the response.
2. If a fallback mechanism is necessary (e.g., due to server load or availability), the user must be explicitly and clearly notified of the substitution before they receive a response from the fallback model.

Trust is paramount in AI development and user interaction. I urge you to address this issue with the seriousness it deserves.

Thank you for your attention to this critical matter.

<img width="638" height="1424" alt="Image" src="https://github.com/user-attachments/assets/335324c1-2b12-4c47-b387-ee687bb23a30" />