This is not a dispute or complaint over normal usage charges. I understand how Codex usage, paid credits, and automatic top-ups work. Over the past five to six months I have used Codex extensively, have regularly exhausted my included usage, and have paid for additional usage without complaint. My account history reflects that I understand and accept normal paid usage.
What happened on July 26 was fundamentally different.
Immediately after switching from GPT-5.5 Medium to GPT-5.6 Sol—following Codex’s recommendation to try the newer model—I assigned a narrowly scoped UI correction involving a fold-state layout issue.
Instead of performing the requested change, the 5.6 sol agent entered what appeared to be a repetitive autonomous analysis loop for approximately eight hours.
Throughout the session it repeatedly reported essentially the same status:
- “Context automatically compacted”
- “No files have been modified.”
- “I’m comparing v0.92 and v0.94…”
- “I’m isolating the posture classifier…”
- “I’m continuing the exact comparison…”
This pattern repeated over and over without meaningful progress.
The task never completed.
No requested UI correction was delivered.
The session ultimately terminated with:
memory allocation of 12912269827 bytes failed
This represents an attempted allocation of approximately 12 GB of memory, after which the application crashed.
During this same runaway execution, repeated automatic credit purchases were charged to my payment method, ultimately resulting in approximately $400 in charges for a task that never completed and never produced the requested work.
In five to six months of using Codex, I have never experienced anything remotely comparable. I sit and do development and I know what to expect byt in all of that time I have never had anything like this happen.
Support has been giving me the runaround since July 27th. I’ve given everything I can. They most recently asked me to go to the thread in the task I’d and submit /feedback. I told them I cannot because when I try opening that thread it literally just site. 12 hours of sitting never loads the thread. I suspect this is because of the size of memory it is trying to load…12gb of data.
Basically I have been charged over $400 in top ups, did not get work results, paid for work that looped for 8 hours and failed. Then support dragged this out and hasn’t responded in days. I not only had all of this happen but also cannot recover any of the work from the thread in the workspace because it is too large and will not open meaning I’ve lost that work and Codex cannot access it within the workspace either.
I have been reading about complaints about 5.6 sol from others. It is, in my opinion, based on research, not stable and needs significant improvements to be used for developers. This is not an isolated comment. I have regressed to 5.5 and am making progress again without the significant added costs of 5.6 sol and runaway processes. Related to this was that shortly after I experienced this Open AI announced lower token costs for 5.6 sol. I wonder why…
Don’t get me wrong Codex has been a game changer for my work but that does not excuse the huge gaping failure and lack of communication from OpenAI which owns Codex…