For the past three days 60-85% of my 5.5 xhigh requests have been serviced by 5.4. I’ve double and triple checked that I’m not setting the wrong model.
My usage is being drained at 5.5 xhigh rates.
I’ve contact support but I just got an AI message back saying “we understand your concern please read the pricing page”.
Has anyone else been experiencing this? You can check at:
chatgptcom/codex/cloud/settings/analytics#usage
If anyone has any information regarding this I’d really appreciate you sharing. It feels bad having my usage draining at 5.5 xhigh rates just to be wrestling with 5.4.
Hi @JS-Prod,
That does look unusual.
Just to narrow it down: was gpt-5.5 xhigh explicitly pinned in the configuration for the affected tasks or was it selected only through the model picker?
Also, were any subagents involved? If so, do you know which models those subagents were configured to use? In some setups, subagents can run with different model and reasoning settings than the primary agent, which could explain why GPT-5.4 appears in the analytics.
well this explains why 5.5 xhigh has been so horrible for hte past month
i think the longer openai refuses to address this issue and stays mute the more damage its doing
ive already moved a lot of my workflow to opus 4.8
its a shame ive been a long time codex and advocate
what bothers me the most is the zero communication and admission that there is serious issues with 5.5
this is not a good look for openai
@LarisaHaster @Agentify.sh
I’ve figured it out! Hopefully this might be able to help someone else!
Do NOT use “Approve for me” as the action approval setting in Codex it spawns a 5.4 sub agent and drains tokens aggressively.
thread_source = subagent
source = {"subagent":{"other":"guardian"}}
model = codex-auto-review
The setting you see selected in the screenshot below is the cause of all the 5.4 requests and rapid usage depletion:
Good thing you figured out that having “Approve for me” enabled was causing a gpt-5.4 subagent to run and drain your usage more quickly.
And thank you for sharing the cause here, in case someone else notices similar thing in their Codex analytics.
