Pro 20x: recurring "at capacity" failures across desktop, CLI and two machines - Sep 21-22

I've read the pinned Anyone Else thread and seen the existing capacity reports. I'm adding a detailed case with timestamps, screenshots and cross-client testing, hoping someone from OpenAI/Codex can help get support case #15497316 reviewed by the appropriate technical team. Support says the case has been escalated and already has my session IDs and diagnostics privately. I don't yet have a confirmed cause or reliable workaround. I'm on Pro 20x ($200/month) , using ChatGPT authentication rather than an API key. Since September 21, Codex has repeatedly stopped in the middle of infrastructure work with: Selected model is at capacity. Please try a different model. It isn't continuously down. Sometimes it reads or edits files, runs a command, then fails again. Brief recoveries haven't translated into reliable operation. September 21: repeated capacity errors interspersed with file reads, edits and command execution. Timeline All times are Taipei time, UTC+8 , in September 2026. September 21, 13:41: The first failure interrupted an ongoing infrastructure task. Errors continued intermittently throughout the day and were still occurring at 21:41 , eight hours later. September 21, separate machine: An OpenClaw agent configured for Astra responded successfully at 17:13-17:14 . A scheduled check then failed at 17:20 with Cause: overloaded . The automation reported recovery at 17:48 , followed by further successful responses. September 22, 14:45: After another period of successful operation, Codex returned the same capacity error. I immediately captured /status in the failing session. What I tested Clients and machines: Failures occurred in Codex within the macOS ChatGPT desktop app, Codex CLI, OpenClaw on my Mac, and OpenClaw on a separate machine using the same ChatGPT/Codex account. OpenClaw also reported overload errors. Models: GPT-6 Astra was the original model. Switching the failing desktop conversation to GPT-5.6 Sol produced the same capacity error. Other checks: Fresh conversations did not resolve it. A new Git worktree briefly appeared to help, but the failure returned there too. Connecting through ProtonVPN with a US exit server did not restore reliable access. ChatGPT Work has also answered during periods when Codex was failing. I don't know whether that difference is significant. Usage during the failures The September 21 Usage screenshot below shows 61% weekly allowance remaining , with the reset in 4 days and 23 hours . September 21: the Usage page showed 61% of the weekly allowance remaining. The /status captured during the September 22, 14:45 failure showed 53% weekly allowance remaining , resetting September 26. These captures don't show a five-hour usage figure, so I'm not claiming that weekly allowance alone rules out every possible limit. I'm asking support to correlate the failed requests with the actual account limits and server-side errors. What I'm asking for A substantial part of my workday went into troubleshooting instead of finishing the task. The main problem is not a delayed answer. It's the agent stopping partway through infrastructure work, repeatedly, without a reliable way to continue. Could someone from OpenAI/Codex help get the failed requests in case #15497316 reviewed and clarify whether this is a usage-limit issue, serving capacity, or another failure? A confirmed explanation and any supported workaround would be useful. If you're experiencing the same error , please add your plan, model/client, timestamps with timezone, and what Usage showed during the failure. Please keep account details, tokens and full logs private. The aim is to give the team concrete cases to compare and help other affected users get their failures investigated too.