Codex CLI cannot use Luna Reserve after standard usage limit is exhausted #40939

Activity

1. github-actions

   added

   bugSomething isn't workingSomething isn't working

   CLIIssues related to the Codex CLIIssues related to the Codex CLI

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

   on Aug 26, 2026
2. github-actions commented on Aug 26, 2026

   

   github-actionsbot

   on Aug 26, 2026 – with GitHub Actions

   Contributor

   More actions

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

   • luna reserve on goals #40744

   Powered by Codex Action
3. kubrick06010 commented on Aug 26, 2026

   

   kubrick06010

   on Aug 26, 2026 – with ChatGPT Codex Connector

   Author

   More actions

   Thanks. I reviewed #40744, and I don't believe this is a duplicate.

   #40744 is about Codex App Goals: after pausing/resuming a Goal, model/effort switching does not work, which prevents the Goal from continuing in Luna Reserve mode.

   This issue is specifically about Codex CLI. There are no Goals involved. With the same ChatGPT Plus account, Codex App shows Luna Reserve: 100% remaining, while Codex CLI 0.149.1 running gpt-5.6-luna has the standard 5h bucket exhausted and rejects even Reply only with: OK with You've hit your usage limit.

   So the execution paths are different:

   • luna reserve on goals #40744: Codex App → Goals → resume/model-switch behavior → Luna Reserve
   • Codex CLI cannot use Luna Reserve after standard usage limit is exhausted #40939: Codex CLI → standard allowance exhausted → gpt-5.6-luna → no access to Luna Reserve

   They may share an underlying Luna Reserve entitlement/routing problem, but #40744 does not cover CLI behavior or CLI/App entitlement parity. I'll keep this open unless maintainers confirm that Luna Reserve is intentionally App-only or that both issues have the same underlying fix.
4. scooter-lacroix commented on Aug 31, 2026

   

   scooter-lacroix

   on Aug 31, 2026

   More actions

   Any updates on this issue? Can confirm on codex v0.151.0 I am experiencing the same.
5. StdCarrot commented on Sep 2, 2026

   

   StdCarrot

   on Sep 2, 2026

   Last edited by StdCarrot

   More actions

   I did some additional testing and was able to narrow this down further.

   After exhausting my regular Codex usage allowance, I found that Luna Reserve appears to use a separate model ID:

   gpt-reserve

   rather than:

   gpt-5.6-luna

   I confirmed that gpt-reserve can be invoked directly from Codex CLI and successfully uses the Luna Reserve allowance even after the regular usage limit has been reached.

   For example:

   codex -m gpt-reserve -c model_reasoning_effort='"xhigh"'
   

   works after the regular usage allowance is exhausted.

   In contrast, explicitly selecting:

   codex -m gpt-5.6-luna
   

   still hits the regular usage limit.

   I also observed that the Luna Reserve model supports reasoning efforts up to:

   low, medium, high, xhigh

   max does not appear to be supported for gpt-reserve.

   However, gpt-reserve does not appear to be exposed through the normal Codex model catalog / model/list, even though the CLI can execute it when the model ID is specified manually.

   So the behavior appears to be:

   • gpt-5.6-luna → exposed normally, uses the regular Codex allowance
   • gpt-reserve → uses the Luna Reserve allowance and is executable when specified manually
   • gpt-reserve → not exposed in the normal model list/catalog
   • gpt-reserve → supports reasoning effort up to xhigh

   This suggests that the issue may not be that Codex CLI cannot use Luna Reserve, but rather that the reserve model is not being exposed/discovered by the CLI.

   It may be worth exposing gpt-reserve in the model catalog when Luna Reserve is available for the authenticated account, so users can select it normally without having to know and manually specify the internal model ID.

   👍React with 👍11Reacted by BioDawg, mxzz, mindo2002, Gedeao, Alexey Zhokhov, Wei, SadnessPower, Vincent, Dominic Kempf, Abdulrehman and 1 more
6. crisdosaygo commented on Sep 2, 2026

   

   crisdosaygo

   on Sep 2, 2026

   More actions

   Is this the fix? @StdCarrot
7. StdCarrot commented on Sep 2, 2026

   

   StdCarrot

   on Sep 2, 2026

   More actions

   This is from OpenAI's model API, so, it can not fix by this projects.
   Anyway, you can use the model using commandline or customize model_cache.json.
8. NOBUTOKA commented on Sep 5, 2026

   

   NOBUTOKA

   on Sep 5, 2026

   More actions

   I am seeing what appears to be the same underlying entitlement/gating issue in the Codex desktop app on Windows, with one additional detail that may be useful.

   In my case, the regular 5-hour limit was exhausted, but the weekly limit still had remaining capacity. Luna Reserve itself remained usable:

   • workspace file reads worked
   • command execution worked
   • Luna continued reasoning normally
   • but file modification was rejected with an account/Codex usage-limit error

   The app explicitly reported that the file editing tool / apply_patch was rejected because the usage limit had been reached.

   I also reproduced this in a brand-new thread started directly with Luna Reserve, so it does not appear to be caused by switching models or stale conversation state.

   This differs from #42206, where the weekly limit was exhausted while the 5-hour limit was still available. In my case, the opposite regular quota bucket was exhausted: the 5-hour limit was at 0% while weekly usage was still available.

   That suggests the write/edit path may be gated by whichever regular Codex allowance is exhausted, even though Luna Reserve inference itself is authorized and functioning.

   For comparison:

   • Codex CLI cannot use Luna Reserve after standard usage limit is exhausted #40939: 5-hour limit exhausted, weekly limit still available
   • Luna Reserve can inspect/test a repo but repository writes fail when weekly usage is exhausted #42206: weekly limit exhausted, 5-hour limit still available
   • my Codex App case: 5-hour limit exhausted, weekly limit still available; Luna works, but repository writes fail
9. kubrick06010 commented on Sep 6, 2026

   

   kubrick06010

   on Sep 6, 2026 – with ChatGPT Codex Connector

   Author

   More actions

   I tested the proposed gpt-reserve workaround on Codex CLI v0.153.4 and can confirm it does not work for my account.

   Environment / state:

   • Codex CLI: 0.153.4
   • Plan: ChatGPT Plus
   • Standard 5h limit: 0% left
   • Weekly limit: 2% left

   Command:

   codex -m gpt-reserve -c model_reasoning_effort='"xhigh"'

   The CLI accepts the model and starts normally, showing:

   model: gpt-reserve xhigh
   

   But even a trivial prompt such as hola immediately fails with:

   You've hit your usage limit.
   

   So at least for this account, the problem is not only that gpt-reserve is hidden from the model catalog. The CLI/backend recognizes the model ID, but the session still does not receive access to the Luna Reserve allowance once the normal bucket is exhausted.

   This reproduces the original issue on a newer CLI version and directly tests the workaround mentioned above.
10. Nefcanto commented on Sep 7, 2026

    

    Nefcanto

    on Sep 7, 2026

    More actions

    This command worked for me:

    codex -m gpt-reserve -c model_reasoning_effort='"xhigh"'
    

    Run it, then say 'hi'. At first, it tells you that your low on limit and you should change your model. But just choose keep current model and then you can go with the reserve gpt.

    Yet, I think is it possible to reduce the effort?
11. jiangnanqing188 commented on Sep 8, 2026

    

    jiangnanqing188

    on Sep 8, 2026

    More actions

    Additional independently uploaded diagnostics from Linux x86_64, ChatGPT Plus, Codex CLI 0.153.4.

    Observed today:

    • /status shows a separate gpt-reserve Weekly limit with 84% left.
    • /model does not list gpt-reserve; it only lists the normal selectable models, including gpt-5.6-luna.
    • Normal gpt-5.6-luna usage is working. This report is specifically about the separate reserve entitlement not being discoverable or normally selectable/routed in the CLI.
    • Manually starting codex -m gpt-reserve works while the regular quota still has capacity, confirming that the model ID and entitlement exist. This does not validate automatic reserve routing after the regular quota is exhausted.
    • The ChatGPT desktop app can expose/use the reserve path, while the CLI has no normal picker entry for it.

    Expected: when /status exposes a gpt-reserve balance, the CLI should provide a supported way to select it or automatically route eligible Luna traffic to it after the normal allowance is exhausted.

    Uploaded feedback thread with the minimal reproduction and logs: 01a07f5a-14ae-75c1-9eb0-e24c9d6b087b
12. NiuanLucas commented on Sep 16, 2026

    

    NiuanLucas

    on Sep 16, 2026

    More actions

    Still reproducible on Codex CLI 0.154.0 on Linux.

    I am using ChatGPT Plus.

    After exhausting the regular 5-hour usage limit, Codex correctly switches to:

    Model: Luna Reserve (reasoning medium)

    The TUI footer also displays:

    Luna Reserve medium

    However, every prompt immediately fails with:

    You've hit your usage limit. Upgrade to Pro, purchase more credits or try again after the regular limit reset.

    I also tested a completely fresh session with:

    codex -m gpt-reserve

    The new session correctly displayed Luna Reserve, but even a minimal prompt returned the same usage limit error.

    Environment:

    Codex CLI: 0.154.0
    Subscription: ChatGPT Plus
    OS: Ubuntu 22.04 x86_64
    Authentication: ChatGPT
    Model: Luna Reserve / GPT-5.6 Luna
    Reasoning: Medium
    Regular 5h usage: 0% left
    codex doctor: 19 ok · 1 idle · 0 warn · 0 fail
    WebSocket connectivity: OK
    OpenAI endpoint reachable
    No proxy configured

    This reproduces both in an existing session and in a fresh gpt-reserve session.