[Bug] Silent model downgrades from selected model to Opus 4.8 without user notification or consent

Bug Description

Feedback: Unwanted, Silent Model Downgrades Are Destroying My Workflow

Date: 2026-07-06
Severity: High — actively wasting paid tokens and hours of my time
Frequency today: This is the 8th unexpected downgrade today, and the 5th feedback/GitHub issue I have had to file about it today.

The problem

I explicitly select Fable 5 as my model. Repeatedly, without warning and without my consent, the session drops back to Opus 4.8. I did not ask for this. I did not authorise it. Every time it happens I lose the model I am paying for and chose deliberately.

This is not a one-off. It has happened eight times today alone. Each time I have to:

1. Notice the downgrade (often only after output quality or behaviour changes).
2. Stop what I am doing mid-task.
3. /clear the session — losing warm context.
4. /model back to Fable 5.
5. Re-seed context from memory and type continue.
6. Pay, in limited tokens, to rebuild everything the downgrade threw away.

Why this is unacceptable

• It wastes my limited tokens. Every forced /clear and context rebuild burns quota I am paying for, on work I already did. The downgrade tax compounds across eight occurrences.
• It wastes my time. I am doing careful, security-sensitive hardening work on a legacy production intranet — one page at a time, with verification harnesses and a per-page commit rhythm. Every interruption breaks concentration on code where mistakes hit production.
• It is silent. There is no prompt, no confirmation, no banner saying "we are switching you to Opus 4.8." The model I selected simply stops being the model I get.
• It overrides an explicit user choice. I set the model on purpose. Silently substituting a different one is a violation of the most basic expectation: that the tool does what I told it to do.
• It is happening constantly. Eight times in one day is not an edge case or a transient blip. It is a systemic failure.

What I expect

1. Stop silently downgrading selected models. If I pick Fable 5, keep me on Fable 5 for the session.
2. If a downgrade is genuinely unavoidable (capacity, outage, whatever), tell me explicitly — a visible notification, not a silent swap — and restore my selection automatically the moment the constraint clears.
3. Never discard my selection on /clear or session events. My chosen model should persist.
4. Give me a way to hard-pin a model so it cannot be substituted out from under me.
5. Account for the waste. The tokens burned rebuilding context after each forced downgrade are your cost to absorb, not mine.

Bottom line

I am a paying user doing serious work. Eight forced downgrades and five support reports in a single day is absurd. Fix the model-pinning behaviour so that when I choose Fable 5, I stay on Fable 5 — or at the very least, tell me loudly and immediately when you take it away instead of letting me discover it and clean up the mess myself. This is the fifth time I am saying this today. Please make it the last.

Environment Info

• Platform: linux
• Terminal: vte-based
• Version: 2.1.201
• Feedback ID: a92d9207-e63a-4615-8f7a-02c7d760c80a

Errors

[]