Codex Rate Limits Discussion Thread

Hi OpenAI / Codex team,

I believe my Codex Pro quota is being metered abnormally. This looks like a quota accounting / bucket assignment / cached-token metering regression, not simply heavy usage.

Current visible status from Codex /status:

• 5h limit: 95% left, resets at 00:33 on June 16, 2026
• Weekly limit: 18% left, resets at 22:16 on June 21, 2026
• In other words, the weekly Codex quota appears to be ~82% used.

The reset timestamp in my usage snapshots matches this weekly window:

• weekly reset: June 21, 2026 at 22:16 Europe/Moscow

I inspected Codex session usage snapshots and found a mismatch between visible quota burn and logged token usage.

Main session investigated:

• Thread/session id: 019e7415-b691-7353-a830-d2231b1c5598
• Thread name: Add branding prototype
• Source: Codex Desktop / VS Code
• Model: gpt-5.5
• Plan type in usage snapshots: pro

Token summary from this session:

• token_count events: 434
• cumulative total_tokens at last snapshot: 55,532,612
• input_tokens: 55,337,946
• cached_input_tokens: 53,562,752
• output_tokens: 194,666
• reasoning_output_tokens: 68,862
• cached input is ~96.8% of input tokens

By date:

• May 29, 2026: about 40.23M total_tokens delta
• June 15, 2026: about 15.10M total_tokens delta in the parent thread

On June 15, this parent thread spawned three subagents:

• 019ecb79-9c9a-7832-a90c-f39f698b115b / Ops
• 019ecb79-b756-75d2-bd09-bd1555113a7f / Main
• 019ecb79-d13d-7070-babf-1770c3fea07c / Vector

The subagent sessions appear to include copied cumulative parent context, so their final total_tokens should not be naively added as separate new usage. Estimated post-fork new token deltas:

• Ops: no post-fork token_count delta visible after 13:30:30Z
• Main: ~1.19M total_tokens
• Vector: ~0.37M total_tokens

So for this investigated parent thread on June 15, the visible logged usage is roughly:

• parent delta: ~15.10M total_tokens
• active subagent deltas: ~1.56M total_tokens
• total estimate: ~16.66M total_tokens

I did not find a literal codex-auto-review source in these snapshots. The visible usage here appears to come from normal agent and subagent execution.

I also inspected the current session that matches the visible /status quota state:

Current session:

• Session id: 019ecc3a-789f-7f51-98fc-773692c51cde
• Model: gpt-5.5

In this current session, over a short diagnostic window:

• first token snapshot: June 15, 2026 17:01:07 UTC
• latest token snapshot: June 15, 2026 17:10:55 UTC
• total_tokens delta: ~3,021,197
• input_tokens delta: ~2,997,763
• cached_input_tokens delta: ~2,716,416
• output_tokens delta: 23,434
• reasoning_output_tokens delta: 8,232

During that same usage window:

• 5h usage moved from 3% used to 7% used
• weekly usage moved from 81% used to 82% used

This is the core concern: a relatively small logged delta of about 3M total_tokens moved the weekly quota by about 1 percentage point, while the account is already showing ~82% weekly usage. If that ratio is representative, it implies an effective weekly allowance of only around ~300M total_tokens, which seems inconsistent with a Pro account and with prior Codex usage expectations.

Please investigate the backend quota accounting for this account, especially:

1. Whether cached input tokens are being charged against included weekly quota at an abnormal or full weight.
2. Whether long-running sessions, compaction, or copied parent context in subagents are being counted repeatedly.
3. Whether subagent/forked sessions are double-counting inherited cumulative context.
4. Whether the account is assigned to the correct Pro quota bucket.
5. Whether GPT-5.5 local usage is being metered with the intended included-limit weighting.
6. Whether the UI is mixing used vs remaining percentages or stale snapshots.
7. Whether there is a recent quota/metering regression around the current weekly reset window.

What I need:

• Account-level review by the Codex quota/accounting team.
• Explanation of why the weekly quota is already ~82% used.
• Restoration or adjustment of the weekly quota if this was incorrectly metered.
• Better usage breakdown in the dashboard: interactive agent vs subagent vs auto-review/background usage, and new usage vs inherited/cached context.

I can provide screenshots of /status and additional aggregate usage numbers if needed.

Also ccusage show this data - 80% of weekly limits on 20x sub lost in 1 day with 500m tokens │ 2026-06-15 │ - gpt-5.5 │ 26,512,423 │ 1,423,718 │ 430,733 │ 539,149,312 │ 567,085,453 │ $444.85 │