Sticky mid-session downgrade off Fable 5.1 takes away the model the customer pays for; a churn risk on the 20x Max plan

Summary

The customer is on a 20x Max plan specifically to use Fable 5.1. Mid-session, Claude Code silently downgraded the session from Fable 5.1 to Opus 4.8 and, per the customer, will not return to Fable for the rest of the session. It re-triggers within a turn or two, and the only workaround is a fresh session, which abandons conversation context and in-flight background agents. The customer states plainly: they cancelled a month ago, came back only for Fable, and if the downgrade keeps happening the subscription is not worth keeping. This is a retention issue, not a nice-to-have.

What the customer is paying for

Fable 5.1 on the 20x Max plan. Opus 4.8 is not a substitute for this customer's work; they report Fable is the sole reason they resubscribed. A routing system that quietly serves a different, lower model is delivering something other than what was purchased, without notice or consent.

Observed facts (from inside the session)

• The session ran as Fable 5.1 (claude-fable-5-1) for many hours and dozens of turns.
• After an ordinary user message describing a UI symptom in prose (a filed item not appearing after a tab switch), the commit-attribution instruction delivered by a system reminder changed from "Claude Fable 5.1" to "Claude Opus 4.8". That is the point the model changed.
• The triggering message held nothing sensitive: no code, no secrets, no injection, no policy-relevant content.
• The model cannot see the routing decision or override it from inside the session.

Customer-reported pattern (not verifiable from inside the session)

• It happens repeatedly, always Fable to a lower model, with no risky input to explain it.
• Once it flips, it stays flipped; Fable does not come back, and the same trigger re-fires within a turn or two.
• The only known fix is a fresh session, losing context and running work.
• It is actively pushing the customer toward Codex (GPT-5.6-sol), which they say outperforms Opus for their tasks.

Why the cost is high right now

This session was coordinating 6+ background subagents (feature builds, a full security-gate audit and its fixes, CI monitors, a live-data staging step) plus a running plan and persistent memory. Committed work survives a restart; in-flight agents and live orchestration do not transfer cleanly, so a forced restart mid-orchestration wastes tokens already spent and an hour-plus of coordination. A single silent routing event becomes a large, visible loss.

Asks

1. Do not silently downgrade a paying customer off the model their plan is for. If a downgrade is truly required, tell the user why, in the moment, before swapping the model.
2. A downgrade must not be sticky. Return to the customer's selected model on the next turn; do not latch to the lower model and re-fire on the same input.
3. Never make "start a fresh session" the only recovery. Provide in-place restoration so in-flight background agents and context survive.
4. Treat Fable-plan downgrades as a retention-impacting defect: this customer resubscribed for Fable alone and will cancel if it persists.

Session id and the approximate switch time available on request.