Image: OpenAI
Codex users on paid ChatGPT plans say GPT-6.1 Terra has been burning through usage much faster since November 5, with one Business-plan report estimating about 15 times more of the five-hour limit per input token than on November 1 through 4. The detailed write-up is GitHub issue 59690 on openai/codex, filed by CoreyHayward. At the time of this writing OpenAI has not posted an on-the-record reply on that issue, on related quota threads, or from OpenAIDevs or Tibo on X, and status.openai.com still shows systems as fully operational.
The reporter said they logged about 21,900 token_count events from Codex sessions covering mid-October through November 5. Model, effort, and service tier stayed on gpt-6.1-terra at low effort and the default tier. Cache hits stayed near 99 percent of input. Mean input per call was smaller on November 5 than earlier in the week, yet the five-hour meter rose about 3.04 percent per Terra call versus about 0.31 percent on November 1 through 4. Per 100,000 input tokens the five-hour cost jumped from about 0.12 percent to about 1.90 percent, which the issue frames as roughly 15.5 times higher. Weekly limit use per call rose about tenfold in the same comparison. GPT-6 Luna calls in the same thread on November 5 still looked normal.
The first expensive Terra burst in that log starts at 8.51 UTC on November 5, which is 2.21 PM IST. The author says the numbers only match OpenAI's published Codex rate card if cached Terra input is now counted at the full input rate even while the client still labels nearly all of it as cached. That is a user hypothesis drawn from their logs, not a confirmed OpenAI change.
Related complaints sit nearby. Issue 59487 describes a weekly Codex allowance falling from about 20 percent remaining to zero in roughly ten minutes on November 4. An OpenAI community thread from a Pro 100 subscriber says remaining allowance fell from 100 percent to 33 percent in about 20 minutes, and that support would not restore usage while the cause stayed unclear. On X, Pro and Business users have tagged Codex staff about sudden empties, resets that still hit limits, and days that burn a full weekly bar. Those posts do not all name Terra or reproduce the 15 times math, but they match the same week of usage anger.
Codex at 0% remaining usage, but it kept working for 8 hours.
— Hai Rozen (@HaiRozen) November 5, 2026
Me whispering to myself, say nothing, don't jinx it.
A day later, Codex burned all 100% in one day.
Me, there must be a bug here. This is madness.
Hey @thsottiaux @ajambrosino, hit a usage limit bug on Codex Pro 100 right after /reset despite 60% weekly limit remaining.
— Jack (@lqjacklee) November 5, 2026
Session 01a1955a-9bba-71c0-bb69-6932fd82cb72
Screenshot attached!
GPT-6.1 Terra launched in late October as OpenAI's cheaper, faster computer-use model. Separately, OpenAI has been moving ChatGPT and Codex toward shared agentic allowances on supported plans, which makes opaque metering harder for users to audit. This Terra report should not be mixed up with the recent ChatGPT Work and Dots permission-errors incident, which OpenAI treated as a different outage and marked resolved.
Until OpenAI confirms a metering bug, a price change for cached Terra input, or a display error, the public case is unpaid user telemetry plus matching complaints. The strongest single dataset remains issue 59690. Paid Codex users who keep Terra in long, highly cached threads are the ones reporting the steepest hits.
Comments