Over the past two days, every Codex task I’ve started on the Windows app has crawled instead of running normally. Jobs that used to finish in a couple of minutes now sit there for ten to twenty, and the longer ones that should wrap in five to ten minutes are dragging on for almost two hours.
The slowdown started late evening on September 26, right after my weekly usage window reset. I kicked off two medium-sized tasks that night, and both behaved the same way. The first one ground away for 48 minutes and eleven seconds before pausing to request clarification. The second one — similar in scope — ran all the way out to one hour and fifty-eight minutes before finally completing.
That alone was strange, but what made it worse was the usage meter. When those two tasks finished, my five-hour window was nearly empty, with only a little over twenty percent left. One of those runs appeared to account for roughly fourteen percent of my total weekly usage by itself. That kind of consumption is completely out of character for what I’ve seen before.
The problem hasn’t let up. This morning I started another task of normal complexity, and it’s still spinning at the time I’m writing this. There’s no error message anywhere — the task just keeps moving, but everything feels ten to twenty times slower than the usual pace.
I’ve tried nothing destructive since the last time this happened a few months ago, which meant wiping my local Codex cache and signing back in. That cleared things up then, but it’s a heavy-handed fix if you’re midway through a project, since it wipes out any cached context. Before I go down that route again, I wanted to check whether anyone else is seeing the same thing right now.
A few questions I’m trying to answer while this keeps running:
Is this tied to GPT-6 Sol Medium specifically, or are other models holding steady? I’m on the latest Codex App version for Windows, ChatGPT Plus plan.
Is there any word from OpenAI about routing or capacity issues in the last day or so? The timing — right after a usage reset — makes me wonder if something got misaligned on their end.
For folks who’ve hit this before: when you cleared your local cache, did that actually fix it, or was it more of a placebo? I’m trying to avoid losing my current context if I can help it. I’m also curious whether the usage meter is reporting accurately. If a task is genuinely taking longer server-side but not actually doing more work, that feels like it should be escalated to OpenAI rather than absorbed by users’ tokens.
I’ve been running Codex daily for development work over the last year, and this kind of broad slowdown across multiple unrelated tasks is unusual. Normally, if one model starts acting up, switching to another smooths things out. That isn’t happening here.
If you’re seeing the same thing — tasks dragging on for tens of minutes with no clear cause, usage dropping fast — drop a note. It might be worth batching feedback together rather than assuming it’s just an isolated setup problem. I’ll update this thread once this current task finishes, or if I end up having to clear my cache and that breaks the logjam.
Next Out-of-order execution for LLM agents, borrowed from CPU design →
All Replies （1） #
Want a live back-and-forth? Join the global AI chat room — login to talk. The slowdown might be due to increased demand during the usage window reset.