Fable 5: repeated crashes, silent auto-switch to 4.6 'due to heavy demand', then context thrashing renders CLI sessions unusable

Summary

Fable 5 sessions in the Claude Code CLI are effectively unusable: repeated crashes, unsolicited auto-switching to another model "due to heavy demand", and aggressive context thrashing after the switch.

Environment

• Claude Code CLI: 2.1.211
• OS: macOS 26.5 (Darwin 25.5.0), Apple Silicon
• Node: v22.21.0
• Configured model: claude-fable-5[1m] (set as "model" in ~/.claude/settings.json)

Symptoms

1. Crashes. Sessions running on Fable 5 crash repeatedly mid-turn.
2. Silent model downgrade. The CLI auto-switches away from Fable 5 to 4.6 with a "due to heavy demand" notice. The configured model is explicit in settings; a capacity fallback that swaps the model out from under a running session is not acceptable behavior for a paid plan, or at minimum needs to be opt-in.
3. Context thrashing. After the downgrade (and sometimes without it), the session enters a compaction/summarization loop that thrashes the context window, losing working state and making long-running sessions useless. The [1m] 1M-context configuration appears to be ignored or mishandled after the fallback, which would explain the thrash if the session is suddenly being squeezed into a smaller window.

Expected

• The model I configured is the model that runs. If capacity forces a fallback, ask, or fail loudly and let me retry; do not switch silently and then repeatedly compact.
• A fallback model change must not invalidate or thrash the existing session context.

Notes

• I am deliberately not attaching session IDs or transcripts.
• Happy to answer clarifying questions about repro conditions in this thread.