{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreibzvb3egwn2ggywx5qm2ruzgzkyokflejhowxd5fcozn2xjhgoqna",
    "uri": "at://did:plc:lk3jfj3zq4k4wxnk474axylu/app.bsky.feed.post/3mkzq3uuju6l2"
  },
  "path": "/t/codex-credits-are-draining-while-idle-not-actively-used/1380250#post_11",
  "publishedAt": "2026-05-04T12:33:55.000Z",
  "site": "https://community.openai.com",
  "tags": [
    "@Foxalabs",
    "@Noah13"
  ],
  "textContent": "Here is a general update for everyone following this.\n\nI received a reply from OpenAI Support. They said they will proceed with a deeper internal review of my account and usage, specifically the session where my Codex credits dropped from 64 to 1 during what was, from my side, normal drafting activity.\n\nThey also said they will check whether a higher-cost model, such as GPT-5.5, was automatically used, and whether background processes such as retries, context handling, or tool-related calls may have contributed to the observed credit usage.\n\nThey explicitly acknowledged that this level of visibility is not currently exposed on the user side, and that investigating it requires internal review.\n\nTheir temporary recommendations were:\n\n  * monitor which model is selected, if visible\n\n  * start a fresh session to isolate behavior\n\n  * avoid prolonged sessions with large context, since that can increase usage\n\n\n\n\nA few clarifications from my side:\n\n@Foxalabs I am not using OpenClaw or any third-party client/wrapper. This is about the official Codex app.\n\nI am fully aware that the dashboard is not instant and that usage reporting can lag. That is not the issue. Even taking delayed reporting into account, the behavior is not consistent with how Codex behaved for me until now.\n\nMy workflow has not meaningfully changed. Same kind of tasks, same kind of sessions, same way of using Codex. What changed is that ordinary procedures that previously used very few credits now appear to consume credits much faster and in a much less predictable way.\n\nAlso, even when I manually leave Codex set to GPT-5.4, which I had been using successfully and was satisfied with, the app often defaults back to GPT-5.5. I would estimate this happens in most cases. That makes usage harder to control because I have to keep checking the selected model every time.\n\nRegarding the Memories point raised by @Noah13: thank you, that was useful. I also passed this information to OpenAI Support.\n\nIn my case, Memories were enabled on at least two machines without me intentionally enabling them. On a third machine they were not enabled. Since turning Memories off, credits no longer seem to be draining “by themselves”, or at least not anywhere near the same level as when Memories were enabled.\n\nOf course, I cannot prove causality from my side. But this gave me a concrete direction to test, and the behavior after disabling Memories does seem different.\n\n@Noah13’s latest description is also very relevant: after a fresh reset, usage dropped significantly even though Codex was apparently not being actively used, only left open while reading a diff. That is exactly the kind of behavior that makes this difficult to understand from the user side. It suggests the issue may not be limited only to visible prompts or explicit user actions.\n\nThis is also why I keep saying this is not just a subjective feeling. Subjectively, yes, it feels like credits are being consumed much faster and less predictably than before. Objectively, my credit balance and actual spending confirm that something has changed.\n\nThe problem is not that Codex costs credits. That is understood. The problem is lack of visibility and predictability. If credits are being consumed by model switching, large context, retries, Memories, background processes, tool-related calls, or idle/open sessions, users need to be able to see that clearly.\n\nAt this point, I have almost developed a conditioned reflex of anxiety when opening Codex, because I no longer know whether a simple task, a long context, an idle window, Memories, automatic model switching, retries, or some background process will start consuming credits in a way I cannot inspect or control.\n\nI hope OpenAI’s review clarifies what happened and that this leads to better transparency for all users, not only for my specific case.\n\nUpdate from today:\n\nThis morning I paid for another 250 credits.\n\nIn less than an hour, I went down to around 150 credits while doing the same type of work and the same processes I had been doing before.\n\nThen I started one more task with 150 credits remaining. That task was not a new complex project or anything unusual — it was basically maintaining/adjusting one existing procedure.\n\nThat single operation took me from 150 credits down to 71 credits.\n\nThis was while using GPT-5.4, not GPT-5.5.\n\nSo again, this is not about me suddenly changing my workflow or doing something dramatically different. These are the same kinds of Codex tasks I was doing before, but the credit consumption is now massively higher and much harder to predict.",
  "title": "Codex credits are draining while idle / not actively used"
}