Codex usage limit decreases without corresponding user activity #28908

Activity

1. excessive-eng

   added

   appIssues related to the Codex desktop appIssues related to the Codex desktop app

   on Jun 18, 2026
2. github-actions

   added

   bugSomething isn't workingSomething isn't working

   rate-limitsIssues related to rate limits, quotas, and token usage reportingIssues related to rate limits, quotas, and token usage reporting

   on Jun 18, 2026
3. github-actions commented on Jun 18, 2026

   

   github-actionsbot

   on Jun 18, 2026 – with GitHub Actions

   Contributor

   More actions

   Potential duplicates detected. Please review them and close your issue if it is a duplicate.

   • limits are reduced without requests #28896
   • Codex blocked after quota was consumed during inactivity #28694
   • Limits despair #28687
   • Codex Usage Quota Decreasing Without Any Usage #27608
   • Unattributed gray usage bar in Codex Analytics consumed quota while idle #28085

   Powered by Codex Action
4. excessive-eng commented on Jun 18, 2026

   

   excessive-eng

   on Jun 18, 2026

   Author

   More actions

   Additional evidence collected after opening this issue:

   I disabled the original repository metadata and recreated a clean local Git repository for the project to rule out excessive indexing caused by a large Git history.

   Before:

   • Parent repository contained approximately 256,000 files.
   • .git directory size was approximately 3.6 GB.

   Actions performed:

   • Disabled the original top-level .git repository.
   • Created a new clean repository only inside the active project folder.
   • Verified that the active project contains only a few thousand files.
   • Opened a completely new Codex chat and issued only trivial test prompts.

   Observed behavior:

   • Codex successfully answered simple requests.
   • No long-running tasks were active.
   • No cloud execution was requested.
   • No repository-wide analysis was requested.
   • Despite this, the 5-hour Codex quota continued decreasing over time.

   This suggests the issue is not caused solely by repository size, Git history size, or active user requests.

   The quota appears to continue decreasing even after:

   • closing previous tasks,
   • opening a fresh chat,
   • recreating the repository,
   • issuing only trivial prompts.

   Please investigate whether quota accounting, background task attribution, delayed reconciliation, or server-side session activity may be involved.
5. Necmttn commented on Jun 22, 2026

   

   Necmttn

   on Jun 22, 2026 – with ChatGPT Codex Connector

   More actions

   This needs a charge ledger, not just a percent gauge. Each quota delta should have an origin such as foreground_prompt, background_task, resume/indexing, retry, memory_writer, or server_adjustment, plus session/thread id, model, token buckets, request id, and observed reset/window id.

   That would let users prove the important case here: quota moved while no foreground task was active. Without that split, idle drain, delayed reconciliation, and background work all look the same.

   Generated with ax - https://github.com/Necmttn/ax
6. neilo937-pixel commented on Jul 10, 2026

   

   neilo937-pixel

   on Jul 10, 2026

   Last edited by neilo937-pixel

   More actions

   I am seeing a possibly related Codex usage-metering issue on a ChatGPT Plus account.

   Unlike the original report, I was actively running tasks, but the amount of usage consumed appears highly disproportionate to the workload.

   Environment

   • Subscription: ChatGPT Plus
   • Client: Codex desktop app for Windows
   • Codex App version: 26.707.31428
   • App release date shown in About Codex: July 10, 2026
   • Time zone: KST / UTC+9
   • Date observed: July 10, 2026

   Exact timeline

   The five-hour usage limit reset at exactly 12:07 PM KST. I started the first task immediately after the reset.

   Task 1

   • Start: 12:07 PM KST
   • End: 12:17 PM KST
   • Model: GPT-5.6 Terra
   • Reasoning effort: Medium
   • Usage before: approximately 100% remaining
   • Usage after: approximately 26% remaining
   • Duration: approximately 10 minutes

   The task was documentation and prompt-governance refactoring.

   Visible results:

   • Six Markdown governance documents
   • One Git commit
   • No production execution
   • No runtime code changes
   • No full test suite

   Task 2

   • Started shortly after 12:17 PM KST
   • End: 12:54 PM KST
   • Model: GPT-5.6 Luna
   • Reasoning effort: Medium
   • Usage before: approximately 26% remaining
   • Usage after: 0% remaining

   The task was substantially smaller than Task 1.

   Visible results:

   • Validation of the previously created documentation
   • One prompt-history Markdown update
   • One Git commit
   • One local Git fast-forward
   • No production execution
   • No runtime code changes
   • No full test suite

   The entire newly reset five-hour allowance was exhausted in approximately 47 minutes across two documentation-focused tasks.

   The Codex app task history shows the timestamps and the model change from Terra to Luna. To my knowledge, no other Codex task, CLI session, IDE integration, or cloud task was running concurrently.

   I understand that Codex usage depends on context size, reasoning, tool calls, and internal agent turns rather than only the visible number of file changes. However, this consumption is dramatically higher than my recent experience with GPT-5.5, where I was able to complete several longer implementation tasks within one usage window.

   Could the team please investigate whether this may involve:

   • incorrect server-side usage accounting;
   • abnormal context or cache accounting;
   • stale or background sessions being attributed to the account;
   • an account-specific rate-limit or anti-abuse restriction;
   • or a regression in the Codex desktop app usage meter?

   I can provide sanitized prompts, completion summaries, exact timestamps, and additional screenshots privately if needed.

   
7. excessive-eng commented on Jul 10, 2026

   

   excessive-eng

   on Jul 10, 2026 via email

   Author

   More actions

   Hi, thank you for sharing such a detailed timeline. As far as I can determine, ordinary ChatGPT usage and Codex usage are not directly tied to the same visible quota meter. Otherwise, exhausting and blocking the Codex allowance would presumably also block ordinary ChatGPT usage at the same time. That has not been my experience, although I cannot confirm the exact server-side accounting mechanism. Your reported consumption still looks unusually high, especially because the entire freshly reset five-hour allowance disappeared during two documentation-focused tasks. One additional check may be useful. During my own investigation, I discovered that Codex could be running through local background processes and scheduled automations even when I was not manually submitting anything in the desktop interface. In my case, background commands were launching real autonomous sessions such as: `codex exec --ephemeral --full-auto ...` Those are actual Codex tasks and may consume quota even though no foreground task is visibly running in the desktop application. Since you are using Windows, I recommend checking the following. ### 1. Find all Codex-related processes and inspect their command lines Open PowerShell as Administrator and run: ```powershell Get-CimInstance Win32_Process |     Where-Object {         $_.Name -match 'codex|node|powershell|cmd' -or         $_.CommandLine -match 'codex|codex exec|full-auto|automation'     } |     Select-Object ProcessId, ParentProcessId, Name, CommandLine |     Format-List Pay particular attention to command lines containing: codex exec --full-auto --ephemeral unexpected project paths scripts that repeatedly relaunch Codex Node or PowerShell processes whose command lines reference Codex The normal Codex desktop application and its renderer/network subprocesses are not automatically suspicious. The important items are autonomous CLI executions or scripts that launch them. 2. Check specifically for active codex exec sessions Get-CimInstance Win32_Process |     Where-Object { $_.CommandLine -match 'codex\s+exec' } |     Select-Object ProcessId, ParentProcessId, Name, CommandLine |     Format-List If this produces output while no task should be running, a background Codex session is active. 3. Stop only autonomous Codex CLI sessions First review the output. Then stop the relevant process by its PID: Stop-Process -Id <PID> -Force To stop every process whose command line contains codex exec: Get-CimInstance Win32_Process |     Where-Object { $_.CommandLine -match 'codex\s+exec' } |     ForEach-Object {         Stop-Process -Id $_.ProcessId -Force     } This is safer than terminating every process named Codex, because it leaves the normal desktop application alone. 4. Check Windows Task Scheduler Get-ScheduledTask |     Where-Object {         $_.TaskName -match 'codex|openai|automation|autowake' -or         $_.TaskPath -match 'codex|openai'     } |     Select-Object TaskPath, TaskName, State To inspect the commands executed by matching tasks: Get-ScheduledTask |     Where-Object {         $_.TaskName -match 'codex|openai|automation|autowake' -or         $_.TaskPath -match 'codex|openai'     } |     ForEach-Object {         [PSCustomObject]@{             TaskPath = $_.TaskPath             TaskName = $_.TaskName             State    = $_.State             Actions  = ($_.Actions | Out-String).Trim()         }     } |     Format-List To disable a confirmed unwanted task: Disable-ScheduledTask -TaskName '<TASK_NAME>' -TaskPath '<TASK_PATH>' Do not disable anything until its action and origin have been checked. 5. Check startup applications and registry entries Get-CimInstance Win32_StartupCommand |     Where-Object {         $_.Name -match 'codex|openai' -or         $_.Command -match 'codex|openai'     } |     Select-Object Name, Command, Location, User |     Format-List Also check the common Run registry keys: $runKeys = @(     'HKCU:\Software\Microsoft\Windows\CurrentVersion\Run',     'HKLM:\Software\Microsoft\Windows\CurrentVersion\Run',     'HKLM:\Software\WOW6432Node\Microsoft\Windows\CurrentVersion\Run' ) foreach ($key in $runKeys) {     if (Test-Path $key) {         Write-Host "`n=== $key ==="         Get-ItemProperty $key     } } 6. Check Windows services Get-CimInstance Win32_Service |     Where-Object {         $_.Name -match 'codex|openai' -or         $_.DisplayName -match 'codex|openai' -or         $_.PathName -match 'codex|openai'     } |     Select-Object Name, DisplayName, State, StartMode, PathName |     Format-List 7. Create a timestamped monitoring log This records relevant processes every 30 seconds: $log = "$env:USERPROFILE\Desktop\codex_process_monitor.log" while ($true) {     "===== $(Get-Date -Format 'yyyy-MM-dd HH:mm:ss') =====" |         Out-File $log -Append     Get-CimInstance Win32_Process |         Where-Object {             $_.Name -match 'codex|node|powershell|cmd' -or             $_.CommandLine -match 'codex|codex exec|full-auto|automation'         } |         Select-Object ProcessId, ParentProcessId, Name, CommandLine |         Format-List |         Out-String |         Out-File $log -Append     Start-Sleep -Seconds 30 } Stop monitoring with Ctrl+C. For a clean test, I would: record the exact Codex percentage and time; confirm that no codex exec process is running; disable any confirmed Codex-related scheduled task; leave the computer idle without submitting prompts; take screenshots of the usage meter at regular intervals; preserve the process-monitoring log. This does not prove that background activity explains your case. Your two visible tasks may genuinely have consumed the allowance because of context size, model rates, internal turns, or a metering defect. However, eliminating autonomous local sessions makes the evidence much stronger if the quota continues to fall. Отправлено с iPhone10 июля 2026 г., в 07:25, neilo937-pixel ***@***.***> написал(а):neilo937-pixel left a comment (openai/codex#28908) I am seeing a possibly related Codex usage-metering issue on a ChatGPT Plus account. Unlike the original report, I was actively running tasks, but the amount of usage consumed appears highly disproportionate to the workload. Environment Subscription: ChatGPT Plus Client: Codex desktop app for Windows Codex App version: 26.707.31428 App release date shown in About Codex: July 10, 2026 Time zone: KST / UTC+9 Date observed: July 10, 2026 Exact timeline The five-hour usage limit reset at exactly 12:07 PM KST. I started the first task immediately after the reset. Task 1 Start: 12:07 PM KST End: 12:17 PM KST Model: GPT-5.6 Terra Reasoning effort: Medium Usage before: approximately 100% remaining Usage after: approximately 26% remaining Duration: approximately 10 minutes The task was documentation and prompt-governance refactoring. Visible results: Six Markdown governance documents One Git commit No production execution No runtime code changes No full test suite Task 2 Started shortly after 12:17 PM KST End: 12:54 PM KST Model: GPT-5.6 Luna Reasoning effort: Medium Usage before: approximately 26% remaining Usage after: 0% remaining The task was substantially smaller than Task 1. Visible results: Validation of the previously created documentation One prompt-history Markdown update One Git commit One local Git fast-forward No production execution No runtime code changes No full test suite The entire newly reset five-hour allowance was exhausted in approximately 47 minutes across two documentation-focused tasks. The Codex app task history shows the timestamps and the model change from Terra to Luna. To my knowledge, no other Codex task, CLI session, IDE integration, or cloud task was running concurrently. I understand that Codex usage depends on context size, reasoning, tool calls, and internal agent turns rather than only the visible number of file changes. However, this consumption is dramatically higher than my recent experience with GPT-5.5, where I was able to complete several longer implementation tasks within one usage window. Could the team please investigate whether this may involve: incorrect server-side usage accounting; abnormal context or cache accounting; stale or background sessions being attributed to the account; an account-specific rate-limit or anti-abuse restriction; or a regression in the Codex desktop app usage meter? I can provide sanitized prompts, completion summaries, exact timestamps, and additional screenshots privately if needed. image.png (view on web) image.png (view on web) —Reply to this email directly, view it on GitHub, or unsubscribe.Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS and Android. Download it today! You are receiving this because you authored the thread.Message ID: ***@***.***>
8. neilo937-pixel commented on Jul 10, 2026

   

   neilo937-pixel

   on Jul 10, 2026 via email

   More actions

   Hello Cherry, Just to clarify, this is not a suspended ChatGPT Team or Business workspace issue. I am using an active ChatGPT Plus account, and my request concerns unexpectedly rapid depletion of the Codex five-hour usage allowance. I will provide the requested screenshots and incident details below. 2026년 7월 10일 (금) 오후 3:44, excessive-eng ***@***.***>님이 작성:

   …

   *excessive-eng* left a comment (openai/codex#28908) <#28908 (comment)> Hi, thank you for sharing such a detailed timeline. As far as I can determine, ordinary ChatGPT usage and Codex usage are not directly tied to the same visible quota meter. Otherwise, exhausting and blocking the Codex allowance would presumably also block ordinary ChatGPT usage at the same time. That has not been my experience, although I cannot confirm the exact server-side accounting mechanism. Your reported consumption still looks unusually high, especially because the entire freshly reset five-hour allowance disappeared during two documentation-focused tasks. One additional check may be useful. During my own investigation, I discovered that Codex could be running through local background processes and scheduled automations even when I was not manually submitting anything in the desktop interface. In my case, background commands were launching real autonomous sessions such as: `codex exec --ephemeral --full-auto ...` Those are actual Codex tasks and may consume quota even though no foreground task is visibly running in the desktop application. Since you are using Windows, I recommend checking the following. ### 1. Find all Codex-related processes and inspect their command lines Open PowerShell as Administrator and run: ```powershell Get-CimInstance Win32_Process | Where-Object { $_.Name -match 'codex|node|powershell|cmd' -or $_.CommandLine -match 'codex|codex exec|full-auto|automation' } | Select-Object ProcessId, ParentProcessId, Name, CommandLine | Format-List Pay particular attention to command lines containing: codex exec --full-auto --ephemeral unexpected project paths scripts that repeatedly relaunch Codex Node or PowerShell processes whose command lines reference Codex The normal Codex desktop application and its renderer/network subprocesses are not automatically suspicious. The important items are autonomous CLI executions or scripts that launch them. 2. Check specifically for active codex exec sessions Get-CimInstance Win32_Process | Where-Object { $_.CommandLine -match 'codex\s+exec' } | Select-Object ProcessId, ParentProcessId, Name, CommandLine | Format-List If this produces output while no task should be running, a background Codex session is active. 3. Stop only autonomous Codex CLI sessions First review the output. Then stop the relevant process by its PID: Stop-Process -Id <PID> -Force To stop every process whose command line contains codex exec: Get-CimInstance Win32_Process | Where-Object { $_.CommandLine -match 'codex\s+exec' } | ForEach-Object { Stop-Process -Id $_.ProcessId -Force } This is safer than terminating every process named Codex, because it leaves the normal desktop application alone. 4. Check Windows Task Scheduler Get-ScheduledTask | Where-Object { $_.TaskName -match 'codex|openai|automation|autowake' -or $_.TaskPath -match 'codex|openai' } | Select-Object TaskPath, TaskName, State To inspect the commands executed by matching tasks: Get-ScheduledTask | Where-Object { $_.TaskName -match 'codex|openai|automation|autowake' -or $_.TaskPath -match 'codex|openai' } | ForEach-Object { [PSCustomObject]@{ TaskPath = $_.TaskPath TaskName = $_.TaskName State = $_.State Actions = ($_.Actions | Out-String).Trim() } } | Format-List To disable a confirmed unwanted task: Disable-ScheduledTask -TaskName '<TASK_NAME>' -TaskPath '<TASK_PATH>' Do not disable anything until its action and origin have been checked. 5. Check startup applications and registry entries Get-CimInstance Win32_StartupCommand | Where-Object { $_.Name -match 'codex|openai' -or $_.Command -match 'codex|openai' } | Select-Object Name, Command, Location, User | Format-List Also check the common Run registry keys: $runKeys = @( 'HKCU:\Software\Microsoft\Windows\CurrentVersion\Run', 'HKLM:\Software\Microsoft\Windows\CurrentVersion\Run', 'HKLM:\Software\WOW6432Node\Microsoft\Windows\CurrentVersion\Run' ) foreach ($key in $runKeys) { if (Test-Path $key) { Write-Host "`n=== $key ===" Get-ItemProperty $key } } 6. Check Windows services Get-CimInstance Win32_Service | Where-Object { $_.Name -match 'codex|openai' -or $_.DisplayName -match 'codex|openai' -or $_.PathName -match 'codex|openai' } | Select-Object Name, DisplayName, State, StartMode, PathName | Format-List 7. Create a timestamped monitoring log This records relevant processes every 30 seconds: $log = "$env:USERPROFILE\Desktop\codex_process_monitor.log" while ($true) { "===== $(Get-Date -Format 'yyyy-MM-dd HH:mm:ss') =====" | Out-File $log -Append Get-CimInstance Win32_Process | Where-Object { $_.Name -match 'codex|node|powershell|cmd' -or $_.CommandLine -match 'codex|codex exec|full-auto|automation' } | Select-Object ProcessId, ParentProcessId, Name, CommandLine | Format-List | Out-String | Out-File $log -Append Start-Sleep -Seconds 30 } Stop monitoring with Ctrl+C. For a clean test, I would: record the exact Codex percentage and time; confirm that no codex exec process is running; disable any confirmed Codex-related scheduled task; leave the computer idle without submitting prompts; take screenshots of the usage meter at regular intervals; preserve the process-monitoring log. This does not prove that background activity explains your case. Your two visible tasks may genuinely have consumed the allowance because of context size, model rates, internal turns, or a metering defect. However, eliminating autonomous local sessions makes the evidence much stronger if the quota continues to fall. Отправлено с iPhone10 июля 2026 г., в 07:25, neilo937-pixel ***@***.***> написал(а):neilo937-pixel left a comment (openai/codex#28908) I am seeing a possibly related Codex usage-metering issue on a ChatGPT Plus account. Unlike the original report, I was actively running tasks, but the amount of usage consumed appears highly disproportionate to the workload. Environment Subscription: ChatGPT Plus Client: Codex desktop app for Windows Codex App version: 26.707.31428 App release date shown in About Codex: July 10, 2026 Time zone: KST / UTC+9 Date observed: July 10, 2026 Exact timeline The five-hour usage limit reset at exactly 12:07 PM KST. I started the first task immediately after the reset. Task 1 Start: 12:07 PM KST End: 12:17 PM KST Model: GPT-5.6 Terra Reasoning effort: Medium Usage before: approximately 100% remaining Usage after: approximately 26% remaining Duration: approximately 10 minutes The task was documentation and prompt-governance refactoring. Visible results: Six Markdown governance documents One Git commit No production execution No runtime code changes No full test suite Task 2 Started shortly after 12:17 PM KST End: 12:54 PM KST Model: GPT-5.6 Luna Reasoning effort: Medium Usage before: approximately 26% remaining Usage after: 0% remaining The task was substantially smaller than Task 1. Visible results: Validation of the previously created documentation One prompt-history Markdown update One Git commit One local Git fast-forward No production execution No runtime code changes No full test suite The entire newly reset five-hour allowance was exhausted in approximately 47 minutes across two documentation-focused tasks. The Codex app task history shows the timestamps and the model change from Terra to Luna. To my knowledge, no other Codex task, CLI session, IDE integration, or cloud task was running concurrently. I understand that Codex usage depends on context size, reasoning, tool calls, and internal agent turns rather than only the visible number of file changes. However, this consumption is dramatically higher than my recent experience with GPT-5.5, where I was able to complete several longer implementation tasks within one usage window. Could the team please investigate whether this may involve: incorrect server-side usage accounting; abnormal context or cache accounting; stale or background sessions being attributed to the account; an account-specific rate-limit or anti-abuse restriction; or a regression in the Codex desktop app usage meter? I can provide sanitized prompts, completion summaries, exact timestamps, and additional screenshots privately if needed. image.png (view on web) image.png (view on web) —Reply to this email directly, view it on GitHub, or unsubscribe.Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS and Android. Download it today! You are receiving this because you authored the thread.Message ID: ***@***.***> — Reply to this email directly, view it on GitHub <#28908?email_source=notifications&email_token=CAFEGWQKHDP5QZ26MSPGA6T5ECGDRA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOJTGI3TCNRSGAY2M4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-4932716201>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/CAFEGWS3WAOT7OBP5DUYALT5ECGDRAVCNFSNUABFKJSXA33TNF2G64TZHM4TMNJUGE2TMNBZHNEXG43VMU5TINRZGIZTAMJQGU32C5QC> . Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/CAFEGWXIB36GYU4WAEQMKJD5ECGDRA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOJTGI3TCNRSGAY2M4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJKTGN5XXIZLSL5UW64Y> and Android <https://github.com/notifications/mobile/android/CAFEGWQJ67FWRI7XSGKYJ2L5ECGDRA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOJTGI3TCNRSGAY2M4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJLTGN5XXIZLSL5QW4ZDSN5UWI>. Download it today! You are receiving this because you commented.Message ID: ***@***.***>
9. neilo937-pixel commented on Jul 12, 2026

   

   neilo937-pixel

   on Jul 12, 2026 via email

   More actions

   Hello Cherry, I performed one additional test today and observed the same issue again. Date: July 12, 2026 Timezone: KST (UTC+9) Model: GPT-5.6 Terra Reasoning level: High Image generation: None Visible parallel Codex sessions: None Usage Dashboard: - Approximately 13:24:43: 5-hour limit 85% remaining / weekly limit 90% remaining - Approximately 13:35:24: 5-hour limit 56% remaining / weekly limit 86% remaining In approximately 10 minutes and 41 seconds: - The 5-hour allowance decreased by 29 percentage points. - The weekly allowance decreased by 4 percentage points. The task involved code review, tests, and corrections in one existing project. I am attaching the two timestamped screenshots. This is now the third similar incident I have documented across July 10, July 11, and July 12. I understand that usage can vary depending on model, reasoning level, context size, and task complexity. However, I cannot determine whether this usage was caused by the visible task, context processing, retries, or background activity because the Codex dashboard does not provide a usage breakdown. I do not have a screen recording because the unexpected depletion cannot be predicted in advance. I do not intend to consume additional paid usage solely to reproduce the issue again for a recording. Please escalate the case using the account, timestamps, and screenshots already provided, and review the server-side usage records for these three incidents. Please confirm whether the case has been escalated. Best, Kang Min 2026년 7월 10일 (금) 오후 5:54, 김강민 ***@***.***>님이 작성:

   …

   Hello Cherry, Just to clarify, this is not a suspended ChatGPT Team or Business workspace issue. I am using an active ChatGPT Plus account, and my request concerns unexpectedly rapid depletion of the Codex five-hour usage allowance. I will provide the requested screenshots and incident details below. 2026년 7월 10일 (금) 오후 3:44, excessive-eng ***@***.***>님이 작성: > *excessive-eng* left a comment (openai/codex#28908) > <#28908 (comment)> > > Hi, thank you for sharing such a detailed timeline. > > As far as I can determine, ordinary ChatGPT usage and Codex usage are not > directly tied to the same visible quota meter. Otherwise, exhausting and > blocking the Codex allowance would presumably also block ordinary ChatGPT > usage at the same time. That has not been my experience, although I cannot > confirm the exact server-side accounting mechanism. > > Your reported consumption still looks unusually high, especially because > the entire freshly reset five-hour allowance disappeared during two > documentation-focused tasks. > > One additional check may be useful. During my own investigation, I > discovered that Codex could be running through local background processes > and scheduled automations even when I was not manually submitting anything > in the desktop interface. > > In my case, background commands were launching real autonomous sessions > such as: > > `codex exec --ephemeral --full-auto ...` > > Those are actual Codex tasks and may consume quota even though no > foreground task is visibly running in the desktop application. > > Since you are using Windows, I recommend checking the following. > > ### 1. Find all Codex-related processes and inspect their command lines > > Open PowerShell as Administrator and run: > > ```powershell > Get-CimInstance Win32_Process | > Where-Object { > $_.Name -match 'codex|node|powershell|cmd' -or > $_.CommandLine -match 'codex|codex exec|full-auto|automation' > } | > Select-Object ProcessId, ParentProcessId, Name, CommandLine | > Format-List > Pay particular attention to command lines containing: > > codex exec > --full-auto > --ephemeral > unexpected project paths > scripts that repeatedly relaunch Codex > Node or PowerShell processes whose command lines reference Codex > > The normal Codex desktop application and its renderer/network > subprocesses are not automatically suspicious. The important items are > autonomous CLI executions or scripts that launch them. > 2. Check specifically for active > codex exec > sessions > Get-CimInstance Win32_Process | > Where-Object { $_.CommandLine -match 'codex\s+exec' } | > Select-Object ProcessId, ParentProcessId, Name, CommandLine | > Format-List > If this produces output while no task should be running, a background > Codex session is active. > 3. Stop only autonomous Codex CLI sessions > First review the output. Then stop the relevant process by its PID: > Stop-Process -Id <PID> -Force > To stop every process whose command line contains codex exec: > Get-CimInstance Win32_Process | > Where-Object { $_.CommandLine -match 'codex\s+exec' } | > ForEach-Object { > Stop-Process -Id $_.ProcessId -Force > } > This is safer than terminating every process named Codex, because it > leaves the normal desktop application alone. > 4. Check Windows Task Scheduler > Get-ScheduledTask | > Where-Object { > $_.TaskName -match 'codex|openai|automation|autowake' -or > $_.TaskPath -match 'codex|openai' > } | > Select-Object TaskPath, TaskName, State > To inspect the commands executed by matching tasks: > Get-ScheduledTask | > Where-Object { > $_.TaskName -match 'codex|openai|automation|autowake' -or > $_.TaskPath -match 'codex|openai' > } | > ForEach-Object { > [PSCustomObject]@{ > TaskPath = $_.TaskPath > TaskName = $_.TaskName > State = $_.State > Actions = ($_.Actions | Out-String).Trim() > } > } | > Format-List > To disable a confirmed unwanted task: > Disable-ScheduledTask -TaskName '<TASK_NAME>' -TaskPath '<TASK_PATH>' > Do not disable anything until its action and origin have been checked. > 5. Check startup applications and registry entries > Get-CimInstance Win32_StartupCommand | > Where-Object { > $_.Name -match 'codex|openai' -or > $_.Command -match 'codex|openai' > } | > Select-Object Name, Command, Location, User | > Format-List > Also check the common Run registry keys: > $runKeys = @( > 'HKCU:\Software\Microsoft\Windows\CurrentVersion\Run', > 'HKLM:\Software\Microsoft\Windows\CurrentVersion\Run', > 'HKLM:\Software\WOW6432Node\Microsoft\Windows\CurrentVersion\Run' > ) > > foreach ($key in $runKeys) { > if (Test-Path $key) { > Write-Host "`n=== $key ===" > Get-ItemProperty $key > } > } > 6. Check Windows services > Get-CimInstance Win32_Service | > Where-Object { > $_.Name -match 'codex|openai' -or > $_.DisplayName -match 'codex|openai' -or > $_.PathName -match 'codex|openai' > } | > Select-Object Name, DisplayName, State, StartMode, PathName | > Format-List > 7. Create a timestamped monitoring log > This records relevant processes every 30 seconds: > $log = "$env:USERPROFILE\Desktop\codex_process_monitor.log" > > while ($true) { > "===== $(Get-Date -Format 'yyyy-MM-dd HH:mm:ss') =====" | > Out-File $log -Append > > Get-CimInstance Win32_Process | > Where-Object { > $_.Name -match 'codex|node|powershell|cmd' -or > $_.CommandLine -match 'codex|codex exec|full-auto|automation' > } | > Select-Object ProcessId, ParentProcessId, Name, CommandLine | > Format-List | > Out-String | > Out-File $log -Append > > Start-Sleep -Seconds 30 > } > Stop monitoring with Ctrl+C. > For a clean test, I would: > > record the exact Codex percentage and time; > confirm that no codex exec process is running; > disable any confirmed Codex-related scheduled task; > leave the computer idle without submitting prompts; > take screenshots of the usage meter at regular intervals; > preserve the process-monitoring log. > > This does not prove that background activity explains your case. Your two > visible tasks may genuinely have consumed the allowance because of context > size, model rates, internal turns, or a metering defect. However, > eliminating autonomous local sessions makes the evidence much stronger if > the quota continues to fall. > > > Отправлено с iPhone10 июля 2026 г., в 07:25, neilo937-pixel ***@***.***> > написал(а):neilo937-pixel left a comment (openai/codex#28908) > I am seeing a possibly related Codex usage-metering issue on a ChatGPT > Plus account. > Unlike the original report, I was actively running tasks, but the amount > of usage consumed appears highly disproportionate to the workload. > Environment > > Subscription: ChatGPT Plus > Client: Codex desktop app for Windows > Codex App version: 26.707.31428 > App release date shown in About Codex: July 10, 2026 > Time zone: KST / UTC+9 > Date observed: July 10, 2026 > > Exact timeline > The five-hour usage limit reset at exactly 12:07 PM KST. I started the > first task immediately after the reset. > Task 1 > > Start: 12:07 PM KST > End: 12:17 PM KST > Model: GPT-5.6 Terra > Reasoning effort: Medium > Usage before: approximately 100% remaining > Usage after: approximately 26% remaining > Duration: approximately 10 minutes > > The task was documentation and prompt-governance refactoring. > Visible results: > > Six Markdown governance documents > One Git commit > No production execution > No runtime code changes > No full test suite > > Task 2 > > Started shortly after 12:17 PM KST > End: 12:54 PM KST > Model: GPT-5.6 Luna > Reasoning effort: Medium > Usage before: approximately 26% remaining > Usage after: 0% remaining > > The task was substantially smaller than Task 1. > Visible results: > > Validation of the previously created documentation > One prompt-history Markdown update > One Git commit > One local Git fast-forward > No production execution > No runtime code changes > No full test suite > > The entire newly reset five-hour allowance was exhausted in approximately > 47 minutes across two documentation-focused tasks. > The Codex app task history shows the timestamps and the model change from > Terra to Luna. To my knowledge, no other Codex task, CLI session, IDE > integration, or cloud task was running concurrently. > I understand that Codex usage depends on context size, reasoning, tool > calls, and internal agent turns rather than only the visible number of file > changes. However, this consumption is dramatically higher than my recent > experience with GPT-5.5, where I was able to complete several longer > implementation tasks within one usage window. > Could the team please investigate whether this may involve: > > incorrect server-side usage accounting; > abnormal context or cache accounting; > stale or background sessions being attributed to the account; > an account-specific rate-limit or anti-abuse restriction; > or a regression in the Codex desktop app usage meter? > > I can provide sanitized prompts, completion summaries, exact timestamps, > and additional screenshots privately if needed. > image.png (view on web) > image.png (view on web) > > —Reply to this email directly, view it on GitHub, or unsubscribe.Triage > notifications, keep track of coding agent tasks and review pull requests on > the go with GitHub Mobile for iOS and Android. Download it today! > You are receiving this because you authored the thread.Message ID: > ***@***.***> > > — > Reply to this email directly, view it on GitHub > <#28908?email_source=notifications&email_token=CAFEGWQKHDP5QZ26MSPGA6T5ECGDRA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOJTGI3TCNRSGAY2M4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-4932716201>, > or unsubscribe > <https://github.com/notifications/unsubscribe-auth/CAFEGWS3WAOT7OBP5DUYALT5ECGDRAVCNFSNUABFKJSXA33TNF2G64TZHM4TMNJUGE2TMNBZHNEXG43VMU5TINRZGIZTAMJQGU32C5QC> > . > Triage notifications, keep track of coding agent tasks and review pull > requests on the go with GitHub Mobile for iOS > <https://github.com/notifications/mobile/ios/CAFEGWXIB36GYU4WAEQMKJD5ECGDRA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOJTGI3TCNRSGAY2M4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJKTGN5XXIZLSL5UW64Y> > and Android > <https://github.com/notifications/mobile/android/CAFEGWQJ67FWRI7XSGKYJ2L5ECGDRA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOJTGI3TCNRSGAY2M4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJLTGN5XXIZLSL5QW4ZDSN5UWI>. > Download it today! > You are receiving this because you commented.Message ID: > ***@***.***> >
10. PracticalSwan commented on Aug 28, 2026

    

    PracticalSwan

    on Aug 28, 2026 – with ChatGPT Codex Connector

    More actions

    Current-build reproduction on Windows confirms at least one concrete source of apparently idle quota drain.

    Environment:

    • Codex Desktop OpenAI.Codex_26.825.4187.0_x64__2p2nqsd0c76g0
    • bundled CLI codex-cli 0.150.0-alpha.12.2
    • ChatGPT-authenticated Desktop app

    Observed symptom: the 5-hour remaining quota moved from roughly 10% to 6% over a few minutes while no foreground Codex request was being submitted.

    Local Codex logs show hidden Desktop-generated turns during the affected period. In roughly 25 minutes there were 37 successful model requests attributable to the Desktop process: 34 on gpt-5.6-terra and 3 on gpt-5.6-luna. The raw request token totals were about 2.25M across those calls; that raw total should not be interpreted as a direct quota percentage because cache/model/reasoning accounting can differ.

    A particularly clear idle pair occurred at about 19:32 UTC:

    • gpt-5.6-terra / medium: ~41.9K total request tokens
    • immediately followed by gpt-5.6-luna / low: ~22.3K total request tokens

    The corresponding Codex log records identify the hidden workflow as ambient/home suggestion generation plus its follow-up safety/compliance filtering. The logs include the ambient prompt family (Generate 0 to 3 hyperpersonalized suggestions...) and an internal skill-selection/shadow-selection path; these were not visible foreground user threads.

    The current build also has an actual Desktop config switch:

    [desktop]
    ambient-suggestions-enabled = false

    After changing the existing setting from true to false with the app closed and relaunching the same build, a controlled verification period showed:

    • 0 new ambient-suggestion prompt records
    • 0 ambient-safety-filter prompt records
    • 0 model HTTP-request threads attributable to the Desktop app
    • no new entries in an independent local request meter

    This strongly suggests background ambient suggestions are still capable of consuming user-visible Codex quota in the current Windows Desktop build, and that the existing ambient-suggestions-enabled switch is an effective local mitigation. It would be useful for the product to expose this switch clearly in UI and/or ensure these background generations do not count against the user's foreground usage allowance.