{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreicz2auwvvz6uk7w4xe4wk4kp4hgebidmfjcztcdurth2hwskkou5a",
    "uri": "at://did:plc:lk3jfj3zq4k4wxnk474axylu/app.bsky.feed.post/3mm5v673buss2"
  },
  "path": "/t/codex-usage-limits-seem-to-be-draining-unusually-fast-since-10-may-2026-even-with-light-gpt-5-3-codex-use-while-the-usage-tracking-ui-also-appears-buggy-or-missing/1380649#post_10",
  "publishedAt": "2026-05-18T21:04:31.000Z",
  "site": "https://community.openai.com",
  "textContent": "One thing that helped me separate real usage drain from dashboard weirdness was keeping a tiny run ledger outside the UI: task id, model, start/end time, approval or elevation count, whether the verifier or acceptance state actually moved, and a usage snapshot before and after each meaningful turn.\n\nIt does not fix the billing bug if there is one, but it makes it much easier to see whether the burn is coming from retry churn, repeated approval loops, or long sessions that never change the outcome. The most useful metric for me stopped being raw tokens and became cost or limit consumed per verified outcome.\n\nIf the usage panel is flaky, having that external receipt also gives support something concrete to compare against. I ended up building MartinLoop around that idea, but even a small CSV or log wrapper around Codex sessions gets most of the diagnostic value.",
  "title": "Codex usage limits seem to be draining unusually fast since 10 May 2026, even with light GPT-5.3 Codex use, while the usage-tracking UI also appears buggy or missing"
}