I downgraded my ChatGPT plan from Pro to Plus on September 16. The morning the $20 charge went through, my weekly Codex allowance read 0%. It stayed there for the better part of a week. Even OpenAI does not know what happens to an account when a reset and a billing date collide. Their own support admits it.

Two days prior to the downgrade, I had redeemed a banked reset. It refilled my weekly window and moved its reset date to September 21. Roughly half of that Pro window was left when the plan changed, and I expected the window to start over at the Plus size. A banked reset is a bonus. I didn’t think I would be penalized for not using all of it. What happened instead was: the date stayed where it was and the meter read 0%. The 5-hour limit sat untouched at 100%.

The first answer I got from AI support was that this is a normal consequence of downgrading. I don’t think it is. I didn’t pay for a banked reset; it was never included. The half I left unused has expired, and I can’t get it back. That is on me. But draining my future usage on top of that puts me in the negative, and a meter that carries over into the following billing cycle sounds suspiciously anti-consumer.

I asked for escalation to human support. The answer was immediate, so I suspect it was generated too; OpenAI leaves that vague. It acknowledged that the public documentation does not say what happens to a weekly window started on Pro when the account moves to Plus midway, and it asked for the time of the reset, the time of the plan change, and my /status output. I sent all three by email. But in the web support chat window, my answer showed up half cut off, and neither the AI nor a human replied.

The weekly limit did reset on its own on September 21. What I lost comes to less than $5 of a $20 plan, and I am not looking for damages. That is what makes the response bizarre. OpenAI should have written it off as a hiccup, handed over a banked reset, and called it a day, or raised the case for a proper review. It did neither.

The fact that banked resets on Codex happen very often, and billing happens every month, makes me wonder if OpenAI bothered to run GPT on their own support documents and Terms and Conditions. Now that Anthropic has introduced a banked reset of its own alongside Opus 5.5, I wonder if Claude will have similar issues in the future.

One possible route to a solution is X, formerly Twitter, where I hope real humans from OpenAI are more present than in customer support. It would be a dramatic move over a $20 plan. But OpenAI hands out the resets and sends the bills regularly. The fact that this isn’t on their radar is alarming. The company should know what happens when the two land in the same week, and it can’t even sort out a simple billing problem.