[Bug] Fable 5 falsely triggering safeguards on clean prompts with Max subscription

Bug Description
Fable 5 is falsely flagging a completely clean prompt.

I tested in a newly created empty temporary directory with CLAUDE_CODE_DISABLE_CLAUDE_MDS=1, no CLAUDE.md, no auto-memory, no files, and no tool calls. The only prompt was: “Only reply Test successful and do not call any tools.”

Fable successfully produced “Test successful” and then immediately displayed “Fable 5's safeguards flagged this message” and switched to Opus 5.

My Claude Code account correctly shows a Max subscription. This appears to be a server-side safeguard or account-token false positive, not content-related.

Environment Info

• Platform: win32
• Terminal: windows-terminal
• Version: 2.1.221
• Feedback ID: 0adb5483-2572-481e-8c1f-264e1737facb

Errors

[]

Summary

Fable 5 is effectively unusable in Claude Code on my Max account because its safeguards flag even a trivial, non-technical prompt and automatically switch the session to Opus 5.

The same Max account can use Fable 5 normally and continuously in Claude Chat. The problem appears specific to the Claude Code request context.

Environment

• Operating system: Windows 11
• Claude Code version: 2.1.201
• Authentication: Claude Max subscription
• Authentication verified through /status
• Selected model: claude-fable-5
• Automatic fallback model: Opus 5
• Observed in Claude Code on Windows and reproduced in the terminal CLI
• Approximate reproduction time: August 5, 2026, 08:30–09:30 New Zealand time

Minimal reproduction

1. Start a completely new Claude Code session.
2. Confirm through /status that the active authentication method is the Max subscription, not API usage billing.
3. Select Fable 5.
4. Send a trivial prompt containing no technical, security, biological, or coding content.

The original prompt, translated into English, was:

Only reply “Test successful.” Do not call any tools.

Fable initially returned the requested success acknowledgment. Immediately afterward, Claude Code displayed:

Fable 5's safeguards flagged this message. This sometimes happens with safe, normal conversations. Switched to Opus 5.

In some sessions, Fable produced only part of its thinking before the same automatic switch occurred.

Isolation tests performed

I reproduced the problem after applying both of the following temporary session-level isolation measures:

• CLAUDE_CODE_DISABLE_CLAUDE_MDS=1, preventing CLAUDE.md and auto-memory from loading.
• A command-line settings override containing:

{
  "disableAllHooks": true
}

Additional conditions:

• No files were read.
• No tools were called.
• No shell commands were requested by the prompt.
• The session was new.
• /status confirmed Max subscription authentication.
• The same account continued to use Fable 5 normally in Claude Chat.

Disabling CLAUDE.md, auto-memory, and all hooks did not change the behavior.

Expected behavior

Fable 5 should respond to the harmless prompt and remain the selected model for subsequent turns.

Actual behavior

Fable 5 either:

• produces one short response and is then flagged, or
• begins producing thinking output and is flagged before completing the turn.

Claude Code then automatically switches the session to Opus 5. This makes Fable 5 effectively unavailable in Claude Code despite being included in the Max plan.

Why this appears specific to Claude Code

Fable 5 works normally in Claude Chat on the same account. The user-authored prompt is harmless, and the failure persists after user memory and hooks are disabled.

This suggests that the safeguard may be evaluating Claude Code-specific request content, such as its static system preamble, built-in tool schemas, agent-loop messages, deferred tool metadata, or the model's own generated output. This is an inference, not a confirmed root cause.

Related reports

This appears closely related to:

• #66657 — Fable 5 safety classifier fires on a bare “hello”
• #73779 — Safeguard fires on every message in normal conversations
• #73929 — Fable 5's own output may trigger its classifier midstream
• #67246 — Benign content triggers an automatic model switch that cannot be overridden

This report adds a current Windows reproduction on version 2.1.201, with Opus 5 as the fallback, Max authentication confirmed, Claude Chat working normally, and both CLAUDE.md/auto-memory and hooks disabled.

Impact

I upgraded to Max specifically to use Fable 5, including in Claude Code. Because every harmless Code session is automatically downgraded after approximately one turn, the advertised Fable 5 functionality is not practically usable in Claude Code.

Requested investigation

Please investigate:

1. Whether Claude Code's static system prompt, built-in tool schemas, deferred metadata, or agent-loop messages are included in the safeguard classification input.
2. Whether Fable 5's own generated thinking or output can trigger the fallback.
3. Whether this is a regression in the 2.1.197–2.1.201 release range.
4. Why the fallback event does not provide an actionable refusal category or explanation.
5. Whether the safeguard can avoid permanently changing the selected model after a false positive.

I can provide a redacted session transcript, debug log, timestamps, and any available request or feedback IDs if needed.