Fable 5 session repeatedly auto-downgraded to Opus 4.8 (suspected safety-classifier false positive on normal deploy work)

Open 💬 1 comment Opened Jul 6, 2026 by AISoloReal

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

Other unexpected behavior

What You Asked Claude to Do

I set the model to Claude Fable 5 in the Claude Code VS Code extension and continued normal software work: editing code and deploying an education web app to my own server (SSH deploy, git push, generating PDFs/graphics).

What Claude Actually Did

The session was repeatedly and automatically downgraded from Fable 5 to Opus 4.8, without me requesting any model change. It keeps happening during ordinary deployment/DevOps steps.

Expected Behavior

The session should stay on the model I selected (Fable 5) for legitimate software development and deployment work, not silently reroute to a lower model.

Files Affected

N/A — this is a model-routing/downgrade issue, not a file-modification issue.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

  1. Set the model to Fable 5 in the Claude Code VS Code extension.
  2. Do normal DevOps work (SSH to a server, run deploy commands, git push, handle deploy keys).
  3. The session gets downgraded to Opus 4.8.

Claude Model

Other

Relevant Conversation

I explicitly selected Fable 5. During normal deployment work (remote SSH deploy of a benign education product), the model was rerouted to Opus 4.8. No dual-use, security, or offensive content was involved.

Impact

High - Significant unwanted changes

Claude Code Version

Max Plan via VS Code extension

Platform

Anthropic API

Additional Context

Suspected cause: a safety/dual-use classifier appears to flag the pattern "remote server access over SSH + deploy keys + running deployment code remotely" as sensitive and reroutes the model. This is a false positive — the work is ordinary deployment of a benign children's education product, with no security or offensive content. Please review classifier calibration for normal DevOps/deployment workflows and forward to the safety/routing team.

View original on GitHub ↗

This issue has 1 comment on GitHub. Read the full discussion on GitHub ↗