I’m on ChatGPT Pro 20x, and my weekly Codex allowance suddenly started draining much faster despite a substantial drop in recorded token usage. I have a Support case open, but I’d like to understand whether anyone else has seen this pattern or can explain the accounting behind it.
I used Codex to compare OpenCodex request logs from my Mac and Linux machine with timestamped allowance readings from CodexBar. The attached chart shows cumulative token usage, split by model, above the weekly allowance consumed. It covers September 8–9, 2026, and all times are Berlin time, CEST.
I also have advanced account security enabled and reviewed the devices signed into my account. I recognized all four: iPhone, iPad, Mac and Linux.
On September 8, between 04:56 and 22:59, the logs recorded 2.180 billion tokens, while the weekly allowance moved by 6 percentage points. Over the following seven hours, ending at 05:58 on September 9, they recorded 472.9 million tokens, while the allowance moved by another 65 percentage points. There was no weekly reset between those readings.
That works out to roughly 50 times more allowance consumed per recorded token in the second period. These token totals include cached input once and exclude separately reported reasoning output. I understand that models, cached input, output, reasoning and service tiers can have different weights. The chart does not establish a 50× billing multiplier, and I wouldn’t expect the two curves to match exactly. What I’m trying to reconcile is the size and timing of the change.
Caching was good tho. The share of input tokens reported as cached actually rose from 97.55% to 98.63% between those periods. The Linux checks also found no sustained cache failure or recorded proxy retry amplification in the six-hour sample examined.
For context, OpenCodex is the local proxy I use to route models through Codex, including models from other providers. This analysis includes only OpenAI-routed requests, and the allowance readings agree with the percentage shown in the ChatGPT app. Those checks give me more confidence in the comparison, although they don’t rule out every client, logging or accounting issue.
One relevant complication is Fast mode. I initially thought it was off overnight, but the logs show priority routing on about 40% of requests during that period, compared with about 10% earlier. Those records describe requested routing; they do not establish which tier was actually billed. I’d like Support to reconcile that alongside the model mix and my Pro 20x entitlement.
I wasn’t using ChatGPT Work during this period. What remains unexplained is the sharp change in allowance consumption relative to the recorded workload. Was the earlier usage unusually discounted or undercounted, was consumption posted with a delay, or was the later usage charged differently?
