Terra High + Chrome integration drains a fresh Plus 5-hour quota from 0% to 100% in under 5 minutes

What version of the Codex App are you using (From “About Codex” dialog)?

26.901.51231

What subscription do you have?

ChatGPT Plus

What platform is your computer?

Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

A single normal Chrome-integrated Codex task using GPT-5.6 Terra with High reasoning exhausted an entire freshly reset ChatGPT Plus 5-hour usage window in 4 minutes 51.948 seconds.

I verified this using the local rollout JSONL rather than relying only on the UI.

The fresh quota reset was recorded as:

2026-09-07T03:58:15.909Z
primary.used_percent = 0.0
weekly.used_percent = 70.0

The same session then reached:

2026-09-07T04:03:07.857Z
primary.used_percent = 100.0
weekly.used_percent = 85.0

So the newly reset 5-hour quota went from 0% used to 100% used in 4m51.948s.

During that interval, local token telemetry showed:

• 30 token-usage/model-continuation updates
• 3,288,054 gross input tokens
• 3,232,640 cached input tokens
• 98.31% aggregate cache hit rate
• only 55,414 uncached input tokens
• 5,022 output tokens
• maximum single input: 122,870 tokens
• 0 requests above 272K input
• 0 recorded cache-write tokens

Prompt caching therefore appears to have been working correctly. The issue is the extremely rapid quota depletion despite almost all retained context being served from cache.

Near the end of the session, one update reported:

122,870 total input
122,496 cached input
374 uncached input
99.70% cache hit

while the 5-hour usage meter increased from 96% to 100%.

The task was interrupted with the UI message that the usage limit had been reached.

What steps can reproduce the bug?

1. Use a ChatGPT Plus account with a fresh Codex 5-hour usage window.
2. Select GPT-5.6 Terra with High reasoning.
3. Start Codex from the Chrome integration / Chrome extension side panel.
4. Give it a normal browser-and-document task that requires several sequential browser operations.

In my case, the task was to:

• find the currently open assignment tab;
• read the assignment requirements;
• open and read an attached worksheet;
• inspect the worksheet editing view;
• perform a few web searches for supporting sources;
• open a teacher-provided YouTube video;
• read the available transcript;
• continue editing the existing document.

Visible tool activity included repeated Chrome tab reads, web searches, YouTube transcript reads, and document interaction.

5. Keep the task running normally without manually starting additional parallel agents.
6. Observe the 5-hour quota.

In my session, the quota had just reset successfully to exactly 0% used.

Session ID:
01a07a02-4232-7ac1-ad89-cc457b355837

Local session metadata:

• originator: codex-chrome-extension-sidepanel
• cli_version: 0.153.4

Observed fresh-window progression included:

03:58:20Z -> 2%
03:58:27Z -> 6%
03:58:34Z -> 10%
03:58:48Z -> 19%
03:59:12Z -> 34%
03:59:44Z -> 49%
04:00:13Z -> 61%
04:00:53Z -> 71%
04:01:21Z -> 80%
04:02:22Z -> 88%
04:03:02Z -> 96%
04:03:07Z -> 100%

The entire fresh 5-hour allowance was therefore consumed in under five minutes.

What is the expected behavior?

A normal single-agent browser/document workflow with approximately 98% prompt-cache reuse should not exhaust an entire fresh ChatGPT Plus 5-hour Codex allowance in under five minutes.

If repeated cached-context replay across Chrome/tool continuations is expected to consume quota at this rate, the product should expose enough usage attribution to make that behavior understandable, such as:

• model continuation count;
• cached vs. uncached input;
• quota consumed per continuation;
• retained context size;
• automatic retries or hidden/background model calls.

I would expect prompt caching to substantially reduce the quota impact of repeatedly carrying forward a mostly unchanged 80K-123K-token context.

Additional information

The rollout data suggests repeated retained-context replay rather than prompt-cache failure.

Across the fresh quota window:

• Gross input: 3,288,054
• Cached input: 3,232,640
• Cache hit rate: 98.31%
• Uncached input: 55,414
• Output: 5,022
• Cache writes: 0
• Maximum single input: 122,870
• Inputs over 272K: 0

The context grew gradually from about 80K to 123K tokens while Chrome/tool continuations were occurring.

Examples:

Input Cached Uncached Cache hit 5h used
80,197 71,552 8,645 89.22% 2%
90,114 87,936 2,178 97.58% 10%
106,011 105,344 667 99.37% 19%
107,672 107,264 408 99.62% 34%
110,235 109,696 539 99.51% 49%
114,808 114,048 760 99.34% 74%
119,482 118,144 1,338 98.88% 88%
122,681 121,088 1,593 98.70% 96%
122,870 122,496 374 99.70% 100%

This does not look like a cache-miss or >272K long-context issue.

The session began shortly before my normal 5-hour reset. The previous window was at 85% used at 03:58:05Z, then the server-reported meter correctly reset to 0% at 03:58:15Z. This makes it possible to clearly isolate the consumption of the new window.

The weekly usage meter also increased from 70% to 85% during the same ~4m52s interval.

I can provide sanitized rollout excerpts if needed.

Start with the local rollout JSONL for session 01a07a02-4232-7ac1-ad89-cc457b355837 and the Chrome extension side-panel continuation path. Compare token-usage updates with the server-reported primary and weekly quota changes, including hidden calls or retries if visible. Done means identifying the accounting cause or documenting why the reported quota increase matches expected usage; no source file or test is named in the issue.