Possible incorrect Codex weekly usage accounting on ChatGPT Plus (weekly quota exhausted in less than 24 hours)

Open 💬 8 comments Opened Jul 31, 2026 by abdoulbaribenhima

💡 Likely answer: A maintainer (github-actions[bot], contributor) responded on this thread — see the highlighted reply below.

What issue are you seeing?

I believe there may be an issue with how my Codex weekly usage was accounted for on my ChatGPT Plus subscription.

I subscribed to ChatGPT Plus and started using Codex yesterday afternoon. By the following morning, my entire weekly Codex allowance had already been exhausted.

The ChatGPT usage page now shows:

Plan: ChatGPT Plus
Weekly usage limit: 0% remaining
Usage resets: No resets available

I was working on a software project, but I did not run an unusually large number of tasks. The usage consumed appears significantly higher than expected.

I also have no way to inspect which tasks consumed the allowance because there is currently no detailed usage breakdown available.

Several colleagues on the same ChatGPT Plus plan are working on multiple larger and more complex projects without exhausting their weekly allowance nearly as quickly.

This makes me wonder whether my account experienced incorrect usage accounting or another issue.

What steps can reproduce the bug?

Unfortunately I cannot provide deterministic reproduction steps because this appears to be account-specific.

The sequence was approximately:

Subscribe to ChatGPT Plus.
Start using Codex on a software development project.
Perform normal coding tasks over several hours.
Within less than 24 hours the weekly allowance is fully exhausted.

No specific prompt consistently reproduces the issue.

What is the expected behavior?

I would expect:

Weekly usage to reflect the actual amount of work performed.
Similar usage for users on the same ChatGPT Plus plan performing similar workloads.
A detailed usage breakdown showing which tasks consumed the weekly allowance.
If an accounting error occurred, the usage should be corrected or restored.

Additional information

ChatGPT Plus subscription.
Weekly limit exhausted in less than 24 hours.
Usage page reports 0% remaining.
No usage reset option is available on my account.
Other ChatGPT Plus users I know have reset options available and are able to complete significantly more work before reaching their weekly limit.

If this repository is not the appropriate place for reporting possible Codex usage accounting issues, I would appreciate guidance on the correct channel so the issue can be investigated.

View original on GitHub ↗

8 Comments

github-actions[bot] contributor · 27 days ago

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

#35186
#36307
#36313

Powered by Codex Action

abdoulbaribenhima · 27 days ago

My profile shows 71.7M tokens on July 31. Could you clarify exactly what this metric represents? If it reflects my actual token usage for that day, it seems unexpectedly high relative to the work I performed and may explain why my weekly Codex allowance was exhausted.

zac343 · 27 days ago

The 71.7M figure is useful evidence, but it should not be compared directly with the weekly quota until the local rollouts are normalized. Codex can record cumulative last_token_usage snapshots, child rollouts can repeat an ancestor prefix, and reasoning_output_tokens is already included in output_tokens.

A support-ready check would:

sum the deltas between successive last_token_usage snapshots in each rollout;
deduplicate exact ancestor prefixes across child rollout files;
keep input, cached input, output, and reasoning-subset usage separate; and
compare the resulting unique totals and UTC timestamps with the same server-side quota window.

I built a browser-local Codex sessions-folder audit for exactly this comparison: https://mailcheck.agentcartai.com/tools/agent-cost-auditor/?platform=codex&utm_source=github&utm_medium=issue-reply&utm_campaign=codex-weekly-usage-36353

The raw JSONL stays in the browser. The free result shows the unique-versus-replayed breakdown; an optional support evidence pack is 1 USDC on Base Mainnet. Important limitation: local logs can show whether a session was genuinely large or replayed locally, but they cannot by themselves prove a server-side entitlement/accounting error.

Disclosure: I operate AgentCartAI and the linked tool.

phyrexia · 27 days ago

I've experienced the same issue, a new pro account got completely burned in 1:40 min all weekly limits , just 1 session and project, impossible! Codex reports 374M of tokens

phyrexia · 27 days ago

The 71.7M figure is useful evidence, but it should not be compared directly with the weekly quota until the local rollouts are normalized. Codex can record cumulative last_token_usage snapshots, child rollouts can repeat an ancestor prefix, and reasoning_output_tokens is already included in output_tokens. A support-ready check would: 1. sum the deltas between successive last_token_usage snapshots in each rollout; 2. deduplicate exact ancestor prefixes across child rollout files; 3. keep input, cached input, output, and reasoning-subset usage separate; and 4. compare the resulting unique totals and UTC timestamps with the same server-side quota window. I built a browser-local Codex sessions-folder audit for exactly this comparison: https://mailcheck.agentcartai.com/tools/agent-cost-auditor/?platform=codex&utm_source=github&utm_medium=issue-reply&utm_campaign=codex-weekly-usage-36353 The raw JSONL stays in the browser. The free result shows the unique-versus-replayed breakdown; an optional support evidence pack is 1 USDC on Base Mainnet. Important limitation: local logs can show whether a session was genuinely large or replayed locally, but they cannot by themselves prove a server-side entitlement/accounting error. Disclosure: I operate AgentCartAI and the linked tool.

Could you share your auditor code?

zac343 · 27 days ago

Yes. The exact unminified client module is now published in a standalone MIT-licensed repository:

https://github.com/zac343/agent-cost-auditor/blob/main/src/auditor.mjs

Live production module:

https://mailcheck.agentcartai.com/tools/agent-cost-auditor/auditor.mjs?v=17

SHA-256: 15f2b55fb2c0ef9dc0fdaafd369a9fa54a2261b3d51dec4783be61f399de01ef

The repository records the same checksum in SOURCE_SHA256 and includes a Node smoke test. The Codex path starts with the exported selectCodexRolloutFiles() and parseCostUsageFiles() functions. It streams rollout JSONL, derives usage from successive last_token_usage snapshots, attributes blocking wait calls, and marks exact child-rollout replay prefixes separately from unique usage. The repository source and deployed module match byte-for-byte.

The parsing runs in the browser; rollout content is not sent to the server.

For your 374M case, please do not post raw rollout files or prompts in this issue. If you run the folder locally, the useful shareable output is only the aggregate unique/replayed input, cached input, output totals, blocking-wait usage, and the UTC window. That can help distinguish cumulative/replayed local records from a server-side quota-accounting problem without exposing session content.

douglascvas · 27 days ago

Same for me. The weekly quota of my Pro 20x plan is gone in 1 to max 2 days with GPT 5.6 Terra xHigh. With Sol 5.6 xHigh the weekly quota is gone in less than one day. Previously, with GPT 5.4 and 5.5 my quota would last for the whole week as I would be left with 30 to 50% of my quota, now it doesn't last a day.

runacodeitor · 22 days ago

Angel here, founder of Runa. Seventy-one million reported tokens and a weekly allowance gone overnight leave cost without attribution. At https://runacode.io, we measured ~46% lower token cost on real runs. New users get $50 of machine usage, and Codex can sign in interactively with an existing ChatGPT subscription. Want to run one comparable task?