No transparent per-feature usage/billing breakdown — quota drains with no attributable local or server-side record #41617

Activity

1. github-actions

   added

   enhancementNew feature or requestNew feature or request

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

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

   on Aug 30, 2026
2. hankwangtw commented on Sep 2, 2026

   

   hankwangtw

   on Sep 2, 2026

   More actions

   Adding a measured data point to this, because it is exactly the gap you describe.

   On my Mac the Chronicle screen-summary pipeline made 6,002 background model calls over the 8 days of logs I still had, and none of them appear in threads.tokens_used or in any rollout file. The local ledger is accurate for user-initiated work; this path simply bypasses it (0 rows whose cwd points at the chronicle capture directory, against 6,030 logged calls that all do).

   Controlled comparison, same machine, app open, no input: 48 minutes with it on → +23% of the 5-hour window; 197 minutes with it off → 0%.

   Counting method, the on/off data and a read-only script that reproduces the numbers are in #42282. Fully agree the fix is attribution, not disclosure — the cost was documented; where it went was not.
3. berkyuo2-cpu commented on Sep 8, 2026

   

   berkyuo2-cpu

   on Sep 8, 2026

   More actions

   Adding a measured Windows data point to the request for usage reconciliation.

   Window: 2026-09-08 04:00–21:11 Europe/Istanbul (UTC+3), i.e. 01:00–18:11 UTC. The cutoff was fixed at the start of the analysis; the subsequent analysis itself is excluded.

   Environment: Codex Desktop package 26.901.6511.0 (Windows package registry), bundled session/runtime version 0.153.4, Windows 10 build 19045 x64, ChatGPT Pro.

   Measured local activity

   51 unique local rollout files: 7 root tasks and 44 subagent threads, with 6,149 distinct usage increments.

   Metric	Tokens
   Input, including cached input	790,370,851
   Cached input (subset of input)	765,014,144
   Uncached input	25,356,707
   Output, including reasoning	4,183,468
   Reasoning output (subset of output)	1,610,602
   Total: input + output	794,554,319

   • Cached-input share: 96.79%. Average input per counted usage increment: 128,536 tokens; input/output ratio approximately 189:1.
   • Root threads: 444,812,048 tokens. Subagents: 349,742,271 tokens (44.02%). One root task plus its children accounted for 48.88% of the total.
   • Peak hour: 11:00–12:00 UTC+3, 128,106,373 tokens, 1,027 increments. Fifteen threads produced usage within one five-minute bucket; that is an activity measure, not proof of simultaneous requests.
   • GPT-6 Astra accounted for 791,530,402 tokens; GPT-5.6 Luna for 3,023,917. Astra effort breakdown: max 546,645,027, ultra 103,371,612, medium 85,970,053, high 55,543,710. Luna used max.
   • 75 compaction records were observed. Their separate allowance impact cannot be derived from these records.

   Account allowance observations

   The previous weekly window was at 99% at 04:04:08. A new reset epoch first appeared at 0% at 04:28:14; a stale old-epoch snapshot also appeared at 04:29:21. The new window reached 95% at 20:55:05, confirmed by a live account usage query at 21:11. The current reset timestamp corresponds to 2026-09-15 04:25:14 UTC+3. Small 1–4-second reset-timestamp variations were not treated as additional resets.

   This is a 95-percentage-point increase in the new window, alongside substantial visible local activity. The trigger for the reset was not established. I cannot determine how much of that account-wide allowance change is attributable to the measured local events versus other surfaces or unlogged work.

   Counting method

   1. Read the local thread index in read-only mode and include older tasks resumed within the window, plus current-day rollouts. A metadata-only coverage check of session/archive files found no additional in-window usage files.
   2. Deduplicate physical file identities, including Windows extended-path aliases. Do not count both paths to one file.
   3. Process token events chronologically within each file and keep the latest pre-window counter as the baseline. Of 6,618 token events, exclude 469 unchanged cumulative snapshots.
   4. 6,098 component-wise cumulative increases exactly matched last_token_usage; 47 first records had total equal to last usage. At 4 counter resets, the new total also exactly equaled last usage, so start a new segment. Arithmetic mismatches: 0.
   5. Sum input + output once. Cached input and reasoning output are subsets. Read the model from turn_context.model, not from rate_limits.limit_id, whose snapshots can refer to different allowance buckets.

   Naively summing every last_token_usage would produce 851,780,170, overstating the reconstructed total by 57,225,851 tokens (~7.20%). This is a log-analysis pitfall; it does not demonstrate duplicate billing.

   Interpretation and request

   The visible workload contains thousands of large-context steps and substantial subagent activity. High cache reuse helps, but repeated cached context still accumulates usage. OpenAI's current pricing documentation explains that model, context, reasoning, tools and caching affect allowance consumption. Raw processed tokens are not directly equivalent to subscription allowance or a monetary bill.

   Please provide a native, time-bounded export that reconciles account allowance with root tasks, subagents, compaction and background/cloud work; distinguishes cached/uncached input and reasoning/output; records request speed/service tier; and documents counter resets, repeated snapshots and the mapping to allowance consumption. The local records here do not expose per-request service tier, so speed-related allowance weighting could not be checked.

   This analysis does not establish overbilling or a regression. Local counter arithmetic is consistent, but there is no independently verifiable account-wide reconciliation from these records. It covers this device's Codex logs, not all ChatGPT web/mobile usage or other devices. Raw prompts, project names/paths, account identifiers, session IDs and credentials are omitted from this public report.