Something feels off with the usage limits

[quecksilber457]

Something feels really broken. My 5h usage limit seems to drop as fast with Luna low as it does with Astra high. Even if it is slower than Astra it burns through my limit way to fast. I asked one question about the memory function and it dropped by 4 % for a 96 words answer and a small search.

[Salavatulin]

I am seeing the same issue, and I was able to quantify it from my local Codex telemetry.

Plan: ChatGPT Plus
Model: gpt-5.6-sol
Reasoning effort: high
Platform: macOS
Codex CLI: currently 0.153.2

I compared local token_count events against the server-reported 5-hour usage percentage.

To reduce contamination from unrelated activity, I restricted the analysis to continuous local Codex activity where consecutive events were no more than 60 seconds apart.

Results:

July 2026:
123 percentage points observed
~1,114,277 raw locally recorded tokens per 1% of the 5-hour quota

August 2026:
87 percentage points observed
~1,430,439 raw tokens per 1%

September 2026:
619 percentage points observed
~214,629 raw tokens per 1%

So September / July = 0.1926, meaning that 1% of the 5-hour allowance now corresponds to about 5.19x fewer locally recorded tokens.

The result is robust to different continuity filters:

60 sec max gap: 5.19x reduction
180 sec: 5.93x reduction
600 sec: 6.01x reduction

September is also consistent across several days on gpt-5.6-sol / high:

Sep 4: ~175k raw tokens / 1%
Sep 5: ~241k / 1%
Sep 6: ~219k / 1%
Sep 7: ~242k / 1%

Late July / early August was generally around 0.8–1.6M tokens / 1%.

I also verified that these quota percentages are coming from server response headers, including:

x-codex-primary-used-percent
x-codex-secondary-used-percent
x-codex-primary-window-minutes: 300
x-codex-secondary-window-minutes: 10080

I understand that raw local token counts are not necessarily identical to OpenAI's internal usage weight, so I am not claiming that the nominal limit was literally reduced by exactly 5–6x.

However, the relationship between comparable GPT-5.6 Sol High local activity and the server-reported quota changed abruptly and by a very large factor.

I have the analysis scripts, rollout data, exact timestamps, and server-header samples available if the Codex team wants them.