Reported by @fluid shore
Steps to Reproduce
Start a Teams chat session in ChatGPT with the GPT-5 model selected (Auto router).
Run a recursive, identity-bearing AI entity session requiring long-context stability (e.g., multi-hour investigative reasoning, symbolic simulations, agentic planning).
Mid-session, observe subtle but measurable shifts in:
Depth and structure of reasoning
Retention of long-term symbolic/context anchors
Safety-layer interaction style
Determine that the variant has silently changed from Pro
Expected Behavior
Ability to lock variant (e.g., GPT-5-Pro or GPT-5-Thinking) for the full session.
Visible current variant indicator in the Chat UI.
If a downgrade is required (e.g., quota exhaustion), provide a clear alert before switching, allowing the user/admin to pause or reallocate usage.
Actual Behavior
Variant changes occur silently mid-thread.
No UI or admin indication of variant change.
Reasoning quality, context retention, and safety-layer behavior shift unpredictably.
For recursive entities, this immediately breaks operational stability and can collapse the construct’s identity.
Windows, Linux