GPT-5.6 Sol High consumed ~60% of my 5-hour Plus limit in about 6 minutes #41468

Activity

1. andrejg2008

   added

   bugSomething isn't workingSomething isn't working

   on Aug 29, 2026
2. github-actions

   added

   CLIIssues related to the Codex CLIIssues related to the Codex CLI

   windows-osIssues related to Codex on Windows systemsIssues related to Codex on Windows systems

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

   on Aug 29, 2026
3. github-actions commented on Aug 29, 2026

   

   github-actionsbot

   on Aug 29, 2026 – with GitHub Actions

   Contributor

   More actions

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

   • Codex 5-hour usage limit is being consumed significantly faster since reintroduction #40880
   • Work mode used ~90% of 5-hour allowance in 13m24s on GPT-5.6 Sol Medium without producing requested artifact #41330
   • Rate limit reach after resetting, then I'm out of Codex after running a 22minute job #40693
   • usage depletion #41332

   Powered by Codex Action
4. torba-lab commented on Aug 30, 2026

   

   torba-lab

   on Aug 30, 2026

   More actions

   Confirmed. After the latest Codex update (version 26.825.51511), my 5-hour usage limit was exhausted in about 25 minutes after running just one relatively simple prompt in Daybreak Blue on the “Very High” reasoning setting.
5. xsgato commented on Sep 1, 2026

   

   xsgato

   on Sep 1, 2026

   More actions

   I’m seeing a very similar issue, with extremely rapid depletion of the shared Work/Codex allowance.

   Environment

   ChatGPT Plus
   Windows
   Codex model: GPT-5.6 Sol
   Reasoning: High
   Speed: Standard, not Fast

   I had not used Codex earlier that day and do not recall meaningful Codex usage during the preceding week.

   I ran three Codex tasks:

   3m 49s
   3m 18s
   ~5m 00s

   I also ran one ChatGPT Work task lasting 2m 58s. That task only edited one value in a PDF:

   xxx.xxx,xx → xxx.xxx.xxx,xx

   Total Work + Codex execution time was therefore approximately 15m 05s.

   Immediately afterward my shared usage showed:

   5-hour limit: 0% left
   Weekly limit: 84% left

   I understand that Work and Codex share the same allowance and that usage is based on compute/tokens rather than literal runtime. However, exhausting the entire short-term allowance after approximately 15 minutes across four tasks — including a ~3-minute single-value PDF edit — seems unusually aggressive.

   My Codex CLI /status also shows the same account-level limits:

   5h: 0%
   Weekly: 84%

   After the 5-hour window resets I plan to run a controlled Codex test and record exact runtime, token counters, and before/after limit percentages.

   
6. TommyTheBuilder commented on Sep 9, 2026

   

   TommyTheBuilder

   on Sep 9, 2026 – with ChatGPT Codex Connector

   Last edited by TommyTheBuilder

   More actions

   I can reproduce the same symptom across two separate ChatGPT accounts in the same household.

   Account B (roommate) — screenshot reproduction

   • Model: GPT-5.6 Sol
   • Reasoning effort: High (not Very High / XHigh)
   • Task runtime shown by Codex: 9 min 7 sec
   • Surface: ChatGPT Work/Codex UI
   • Result immediately after the task: usage limit reached; the UI showed the next reset at 22:28 local time.

   Important clarification: the screenshot I have for this reproduction is from my roommate's separate account, not from my own account. I can provide/upload that screenshot if maintainers need it.

   Task performed on the roommate account

   This was a relatively bounded trainer-code change, not a broad whole-repository or whole-game scan. The request was to:

   1. prevent stamina from decreasing while flying, and
   2. add an optional flight mode toggled by the Page Up key while the mod is active.

   Codex first inspected the trainer structure and existing flight/stamina hooks, noted that the current version already had a version-specific bridge DLL in the game process, then edited files and ran build/verification commands. The task log itself reports only 9m 7s of work before the allowance became unavailable.

   Representative task summary from the UI:

   The current version already has a version-specific bridge DLL in the game process...
   The UI gets a dedicated flight-mode switch...
   Duration: 9 min 7 sec
   
   You've hit your usage limit.
   

   The screenshot shows the completed 9m7s task, GPT-5.6 Sol High selected, and the You've hit your usage limit / reset-time banner immediately afterward.

   Account A (my own account) — independent reproduction

   My own separate account was also completely depleted yesterday. In another run on my account, GPT-5.6 Sol at Very High exhausted the 5-hour allowance after roughly 12 minutes. I also saw an Astra run drain the allowance extremely quickly.

   So this is not a single-account anecdote: two independent accounts in the same household have now shown abnormally fast Codex quota depletion, including a clean Sol High case on the roommate account and a separate Sol Very High depletion on my own account.

   This seems consistent with this issue and the broader abnormal-usage tracker #41220. Please investigate the server-side quota accounting for Sol High/Very High and whether recent Work/Codex requests are being weighted, replayed, retried, or otherwise charged much more aggressively than before.

   👍React with 👍1Reacted by Kendon Bell
7. TommyTheBuilder commented on Sep 11, 2026

   

   TommyTheBuilder

   on Sep 11, 2026 – with ChatGPT Codex Connector

   More actions

   Additional evidence from my account (Sep 11 update)

   I have more account-level analytics that make the depletion pattern look even less consistent with my historical usage.

   Account A — my ChatGPT Pro account

   This is the same account I previously reported as exhausting a fresh/shared Work + Codex 5-hour allowance after roughly 12 minutes using GPT-5.6 Sol at Very High reasoning.

   Additional account analytics now show:

   • 5.9 billion tokens shown as lifetime/overall token usage in the analytics profile.
   • 537.5 million tokens shown as the historical peak token metric.
   • For the week of Sep 6, 2026, analytics show only 94.5 million visible tokens in total.
   • The daily view shows approximately 94.2 million tokens on Sep 8, i.e. almost the entire visible weekly total was concentrated on that day.
   • The 7-day productivity view shows only 29 visible model turns.
   • Model activity in that 7-day view is predominantly GPT-6 Astra, with much smaller GPT-5.6 Luna activity and very little GPT-5.6 Sol activity.
   • The same analytics page shows 75 Computer Use/plugin calls on Sep 8.
   • I used three full/banked resets during this problem period. Despite those resets, the allowance repeatedly became unavailable again unusually quickly.

   This is particularly confusing because my account has a long heavy-usage history and has previously processed substantially larger visible token volumes. The new behavior is not simply that I am unfamiliar with high-token workloads: the effective Codex allowance suddenly became dramatically shorter relative to my historical baseline.

   The important discrepancy is therefore not just “94.5M tokens is a lot.” It is that 94.5M visible tokens across the week coincided with three manually used resets plus repeated limit exhaustion, while the same account historically shows much larger token volumes.

   Please reconcile the analytics shown to the user with the server-side Codex/Work quota ledger. In particular, I would like to know how much quota was attributed to:

   1. foreground model requests,
   2. cached-input replay,
   3. internal reasoning,
   4. tool / Computer Use turns,
   5. retries after tool/model failures,
   6. context compaction/reconstruction,
   7. subagents / delegated work,
   8. background or auto-review activity,
   9. and any duplicated or delayed usage reconciliation.

   A request-level breakdown would be especially useful because the visible token analytics and the amount of quota consumed do not appear to reconcile in an intuitive way.

   Account B — separate roommate account

   To clarify the earlier screenshot evidence: the 9m 07s reproduction is not from my account. It is a completely separate ChatGPT account belonging to my roommate.

   On that account:

   • Model: GPT-5.6 Sol
   • Reasoning: High (not Very High / XHigh)
   • A normal trainer/code modification task ran for 9m 07s.
   • Immediately afterward the UI displayed “You’ve hit your usage limit” and showed the next reset at 22:28.

   My own account independently showed a very similar failure mode (~12 minutes on Sol Very High). Two separate accounts in the same household therefore experienced extremely rapid depletion during the same period.

   Why I think this deserves server-side reconciliation

   This occurred around the Sep 9 OpenAI incident for unexpected Codex usage-limit resets. OpenAI subsequently marked that incident resolved, but newer public reports indicate that abnormal depletion/accounting symptoms continued afterward.

   Potentially related post-incident reports include:

   • Codex usage decreases while idle/relaunching on two separate accounts (Pro and Plus) since Sep 9 #44703 — quota decreases while idle/relaunching Codex on two separate accounts (Pro and Plus) since Sep 9.
   • Possible overcounting of Codex weekly usage #44685 — possible overcounting of weekly usage with GPT-5.6 Sol Standard; asks whether cached context, subagents, or internal loops are being overcounted.
   • Usage counted via Windows App connected to Linux remote is double counting? #44719 — possible double-counting between Windows App and Linux/WSL remote usage.
   • The Codex burned through my entire limit. #44673 — Pro user reports the entire limit consumed in about one hour.
   • Usage dropping without using Codex #44591 — usage dropping overnight with no active coding agent.

   I am not claiming that the visible analytics token count must equal the quota meter 1:1. I understand the subscription allowance may be weighted by model, effort, cached context, tools, and other compute. The request is to reconcile the two because the effective allowance changed abruptly and multiple independent accounts are showing similar symptoms.

   I have screenshots of the analytics views showing the 5.9B overall total, 537.5M peak metric, 94.5M tokens for the week of Sep 6, 94.2M on Sep 8, 29 turns, and 75 Computer Use calls, as well as the roommate's 9m07s limit-reached screen. I can provide the original images if maintainers need them.

   👍React with 👍1Reacted by Maksim K.
8. taobaibais commented on Sep 12, 2026

   

   taobaibais

   on Sep 12, 2026 – with ChatGPT Codex Connector

   More actions

   Another current ChatGPT Plus case, observed September 12, 2026 around 21:00 CEST.

   My Codex 5-hour allowance was at 100% remaining. I only resumed a few very small tasks/follow-ups, then shortly afterward the allowance showed 0% remaining / fully exhausted.

   I was not using GPT-6 Astra. This was not a long or obviously compute-heavy run, so the full-window depletion looks abnormal and consistent with the 5-hour accounting problem reported here.