Codex: overspending of limits

What version of Codex is running?

Web version, at chatgpt.com/codex

What subscription do you have?

ChatGPT Plus

Which model were you using?

Codex

What platform is your computer?

Darwin 22.6.0 arm64 arm

What issue are you seeing?

Since October 30, credit usage has increased dramatically, and they're depleted very quickly.
Simple tasks involving writing a few lines of code eat up a significant portion of the daily and weekly limits.

Even if this is a new policy, and the limits have indeed been lowered (which no one announced; on October 30, the only news was that if the limits are exhausted, credits can now be purchased; the limits themselves remain the same for Plus subscription users), it's impossible to purchase these credits. A request to https://chatgpt.com/backend-api/payments/checkout attempting to purchase 1,000 credits for $40 fails; the server returns a 500 Internal Server Error, which is visible in the browser console.

It turns out that the weekly limit is depleted in a day of active work, and after that, there's nothing to be done except wait for the limits to reset, which won't happen for another week.
Work is at a standstill, and we have to urgently switch to other agents and other systems, which spoils the quality of the code and slows down development.

What steps can reproduce the bug?

Try creating a subscription account plus a few simple tasks, write a few messages, and observe the limits being overdrawn

What is the expected behavior?

normal consumption of limits, which was before October 30

Additional information

Perhaps the token counter has broken, making each task significantly more expensive, or the limits have been lowered for Plus users (which hasn't been reported, so it's unlikely), or perhaps the Codex model is incorrect.

Recommended actions: roll back the site and Codex service to the state before October 30, reset the limits for all users (or recalculate them using the correct numbers), send a notification to all users who have used Codex since October 30, possibly by email, notifying them that the limits are now working correctly, offering an apology. Then, separately refine the new token purchasing system without affecting the functionality of the production site.

How to prevent this in the future: restore the ability to communicate with real support through one method or another on help.openai.com. Currently, there is only an AI assistant available, which can only provide consultations but cannot create tickets or requests. If you ask him to escalate the issue to support, he'll suggest doing so on help.openai.com, but that's simply not possible there. This certainly saves money on support, but it cuts the development team off from real users and their feedback, since the average user simply can't find where to complain or report a problem, so they have to go here (I'm not sure anyone from Codex even looks at issues here, unfortunately).