I am a paying Google AI Pro subscriber using Gemini for heavy quantitative software engineering and Python development inside Custom Gems on Windows 10. Ever since the October 13 model rollout—and escalating into a total server collapse between November 10 and November 12, 2026—Gemini 4 Argon (Thinking: High) has become completely unusable.
Here is the exact breakdown of what is happening right now (screenshots attached):
- 10–15 Minute Thinking Hang & “An error occurred” in Coding Sessions: Whenever a coding thread in a Custom Gem reaches around 15–20 turns (or even 3–4 turns over the last 48 hours), Gemini 4 Argon (High) hangs for 10 to 15 minutes straight, triggers an unrequested “Show code <>” execution block, truncates the code mid-line, and crashes with “An error occurred”. Once this banner appears, the entire thread is permanently bricked. Clicking “Try again”, editing the previous prompt, refreshing, changing IPs, or attaching prompts as .txt files all fail with the exact same error.
- November 12 Escalation — 5 Minutes for “Hi” & Silent Chat History Rollbacks: As of today (Nov 12), the backend doesn’t even reach the thinking phase on existing coding threads. When I ask Argon (High) to complete my truncated code, it just spins on the three-dot loading animation for several minutes, silently drops the request without even throwing an error message, and abruptly rolls the entire chat window back to the last saved message from days ago! Even in a brand-new chat, sending a simple “hi” takes 5 minutes to respond.
- Broken & Inconsistent Gemini 4 Carbon Rollout (Android vs. Windows): For the exact same Google AI Pro account, Gemini 4 Carbon is unlocked on the Android app (where Argon was completely removed, though Carbon on mobile doesn’t even know which model it is when asked!), while on Windows (Web & Desktop App), Carbon is locked behind the Ultra tier and Pro users are stuck with a completely broken Argon model.
- Massive Regression vs. Gemini 3.1 Pro: Prior to October 13, Gemini 3.1 Pro handled massive, hundred-turn coding threads flawlessly like clockwork. Why push out new models every few weeks when the backend infrastructure lacks the TPU capacity and timeout configuration to run existing models?
Please acknowledge this issue, fix the server-side gateway timeouts and session rollbacks on Gemini 4 Argon (High), and resolve the Pro tier mismatch on Windows immediately.
