Codex App consumes ~20-23% of 5-hour quota for a single small UI task #46689

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

26.908.40401

What subscription do you have?

ChatGPT Plus

What platform is your computer?

Windows 11 Pro 25H2 Microsoft Windows NT 10.0.26200.0 x64

What issue are you seeing?

Terra Medium consumed approximately 20% of my 5-hour usage quota while completing a relatively small, well-scoped UI task.

Environment

• Model: Terra Medium
• Subscription: ChatGPT Plus
• Codex App version: 26.908.40401
• OS: Windows 11 Pro 25H2
• Session: new chat
• Context compaction: none
• Task duration: 7 min 33 sec
• 5-hour quota consumed: approximately 20%

Prompt

The task was:

docs/gui-redesign/lesson-workspace/prototype v4 defines a limited ws-paper area for static mode.

Task 1: reduce the board workspace from the top by 50px so it starts below board-v2-workspace-bar.

Task 2: make the static-mode paper behave like the prototype.

Note: opening the Ledge must not resize the paper.

Result

The task completed successfully.

The implementation:

• moved the board render area 50px below the workspace bar;
• added a centered 1920x1080 static paper with free space around it;
• ensured opening the Ledge does not resize the paper;
• updated the related contract documentation;
• added a regression test.

Final diff:

6 files changed, +146 / -33

Checks passed:

• pytest regression test: 2 passed
• node --check BoardApp.js: passed
• node --check LessonShell.js: passed

Issue

For a task of this size, consuming approximately 20% of the entire 5-hour quota in only 7 minutes 33 seconds seems unexpectedly high.

This was a new chat with no accumulated conversation history and no context compaction.

At this rate, roughly five comparable small UI tasks could exhaust the entire 5-hour quota.

Could you please confirm whether this usage is expected for Terra Medium, or whether there may be an issue with quota accounting or repeated context processing?

What steps can reproduce the bug?

I observed unexpectedly high 5-hour quota consumption in two independent runs of essentially the same small UI task.

Run 1 — Terra Medium

1. Started a new Codex App chat using Terra Medium.
2. Asked Codex to:
   • move the board workspace below board-v2-workspace-bar;
   • implement the prototype's centered static ws-paper behavior;
   • prevent the Ledge from affecting paper sizing.
3. The run took 7 min 33 sec.
4. Resulting diff:
   • 6 files changed
   • +146 / -33
5. Context window at completion:
   • 194k / 258k (75%)
6. Approximately 20% of the 5-hour quota was consumed.

The resulting implementation was not correct. It introduced a regression where board wallpapers disappeared and did not fully achieve the requested behavior.

I completely reverted these changes.

Run 2 — Sol Medium

After the rollback, I started another run using Sol Medium.

The prompt contained MORE context than the Terra prompt because I additionally described:

• the expected prototype behavior;
• that the paper must resize when the Rail opens/closes;
• that the Ledge must NOT resize it;
• the implementation approach attempted by the previous model;
• the regression caused by that implementation.

Sol Medium completed the task correctly.

At completion:

• Context window: 162k / 258k (63%)
• 5-hour quota: decreased from 27% remaining to 4% remaining
• Quota consumed by this run: approximately 23%

Despite receiving a larger prompt and successfully solving the task, Sol used less context window than Terra (162k vs 194k), while still consuming approximately 23% of the entire 5-hour quota.

Observation

Two relatively small UI implementation runs consumed approximately:

• Terra Medium: ~20%
• Sol Medium: ~23%

Together, two tasks consumed roughly 43 percentage points of the 5-hour allowance.

This seems unexpectedly high for tasks of this scope.

What is the expected behavior?

I expect a small, well-scoped UI implementation task to consume a substantially smaller fraction of the 5-hour quota.

In particular, quota consumption should be reasonably related to the actual amount of model/tool work performed.

The Sol run is especially notable:

• it used 162k / 258k (63%) of the context window;
• it completed the task successfully;
• but it consumed approximately 23% of the entire 5-hour quota.

The Terra run similarly consumed approximately 20%.

I would like to know whether this level of quota consumption is expected under the current Codex limits, or whether there may be a regression in quota accounting or model/tool-call usage.

Additional information

The Terra run was a new chat and no context compaction occurred.

The Sol run was performed after completely reverting the Terra changes, so both runs started from effectively the same repository state.

An interesting difference is that the Sol prompt contained more information than the Terra prompt, but final context usage was lower:

• Terra Medium: 194k / 258k (75%)
• Sol Medium: 162k / 258k (63%)

Nevertheless, both runs consumed around one fifth of the full 5-hour quota.

I cannot determine whether the underlying allowance has changed, so I am reporting only the observed usage values.