FULL STORY

Codex Quota Saga: From Secret Cuts to Refunds

OpenAI Codex users repeatedly complained of quotas burning abnormally fast, forcing OpenAI to admit bugs and fully reset paid allowances—yet disputes over the new Astra model reignited the controversy.

2026-08-12 ~ 2026-09-27 · 10 episodes · 47 posts

Episode 1 · OpenAI Codex Active Users Exceed 15 Million with Reset Quotas (2026-08-12, 4 posts)

OpenAI Codex's active users have surged to over 15 million in less than six months. To celebrate this milestone, the company has reset and significantly increased usage quotas for users, resuming an activity that was previously paused after hitting 10 million users.

Episode 2 · OpenAI Codex Users Report Quotas Drained Rapidly, Quality Complaints Mount (2026-08-19, 4 posts)

Multiple OpenAI Codex Pro users report that their quotas are being consumed abnormally fast, with some hitting limits in 20 minutes or burning 80% of weekly quota in a day, while heavy users found the $200 Pro plan lasts only about 12 hours. Complaints about declining code quality are also mounting.

Episode 3 · OpenAI Codex Hits 20 Million Active Users, Resets Usage Credits (2026-08-21, 4 posts)

OpenAI announced Codex surpassed 20 million weekly active users, a 20x increase in six months, and reset banked credits for all Codex and ChatGPT Work users while apologizing and investigating abnormal usage consumption.

Episode 4 · OpenAI Admits Codex Cache Hit Rate Dropped After Users Report Quota Anomalies, Promises Compensation (2026-08-22, 7 posts)

On Aug 22, OpenAI Codex users reported a surge of abnormal quota consumption. Product lead Tibo Inoue's initial denial was rebutted by community evidence, and the incident closed for now with OpenAI acknowledging a drop in cache hit rates and promising compensation.

Confirmed

  • MaziyarPanahi reported burning 60% of his weekly quota in a single day with token usage identical to normal; Codex team members he tagged said they found no issue. RileyRalmuto, who has long built the same app with Codex, measured at least 20% higher consumption for the same workload, suspecting either a more token-hungry model or a quiet quota cut.
  • Inoue later partially acknowledged the problem, saying cache hit rates for some users this week fell below previously stable levels, accelerating consumption; the team is investigating with an update expected the next day, and compensation will be offered.
  • Inoue's initial denial, which included accusing users of fraud, was rebutted by community evidence and flagged by X's Community Notes.

Unconfirmed

  • The root cause (model behavior vs. caching infrastructure) awaits the official investigation.
  • RileyRalmuto felt consumption slightly improved afterward but could not confirm a fix.
  • The scope and form of compensation remain unclear.

Why it matters

  • The episode shows how X's Community Notes can instantly correct vendor claims; executive denial holds little ground against user evidence.
  • For developers relying on Codex, quota consumption directly affects cost and availability—fluctuations in infrastructure parameters like cache hit rates materially hurt users, and transparent communication beats denial.

Episode 5 · OpenAI Users Complain of Frequent Usage Limits Without Reset (2026-08-22, 2 posts)

Multiple OpenAI users report hitting usage limits prematurely, with no reset even after upgrading to the $100 plan, raising questions about OpenAI's quota management and service stability.

Episode 6 · OpenAI Admits Three Codex Bugs, Resets All Paid Usage (2026-08-23, 6 posts)

OpenAI Codex lead Tibo Sottiaux posted a long thread admitting that three issues caused Codex users' quotas to drain abnormally fast, and announced a full usage reset for all paid subscriptions as compensation. The issue manifested as rate-limit problems on Codex Remote (including the iOS ChatGPT app). A tiger team is already fixing the bugs and has found a new method to significantly boost efficiency.

Confirmed

  • Three root causes: inefficient image handling in long sessions (wasted compute in the image compression pipeline), high P95 usage load from the Computer History feature, and excessive compute consumed by automatic session title generation
  • A tiger team is fixing these issues, with fixes shipping the next day
  • The team also discovered a new method for significant efficiency gains, planned for the following week
  • Codex usage for all paid subscriptions will be fully reset at 14:00 PST the next day

Why it matters

  • Fast-quota drain directly reduced paid users' usable capacity; the full reset amounts to compensation for subscribers
  • If the promised efficiency improvements land, they could improve Codex's overall resource usage and experience

Episode 7 · OpenAI Restores 5-Hour Codex and Work Usage Limits for ChatGPT Plus (2026-08-25, 4 posts)

OpenAI is reinstating 5-hour usage limits on Codex and Work for ChatGPT Plus users starting August 25, ending the unlimited build mode to smooth compute load and protect weekly quotas, while Pro users are treated differently.

Episode 8 · OpenAI Codex Users Report Quotas Draining Abnormally Fast (2026-09-05, 5 posts)

From September 5 to 7, multiple paying OpenAI Codex users reported that their quotas were draining abnormally fast, spanning the $200 Pro tier and the Pro 5x plan. They suspect OpenAI quietly changed the quota calculation, but there has been no official confirmation.

Confirmed

  • Blogger xiaohu (09-05) reported that the Codex quota on the $200 ChatGPT Pro tier, which previously lasted days, now runs out in half a day, and asked other users about similar experiences.
  • Developer danielmac8 (09-07) measured a task from a video that ran 19 minutes and consumed 5% of the Pro 5x weekly limit, implying the weekly quota is very tight; he also noticed the Codex UI lagging noticeably during the task, suspecting task load.
  • User tawnniee complained about frequently hitting usage limits on Codex Pro, something never experienced before.
  • User chongdashu (09-06) reported that since Codex's usage reset, weekly quota drains noticeably faster: previously running 5.6 Sol with fast tier plus xhigh or even max reasoning intensity did not burn through quota this quickly.

Unconfirmed

  • The exact cause—billing calculation changes, usage statistics changes, or something else—remains uncertain among all posters.
  • A comment on tawnniee's post mentions "GPT 6 Astra is token efficient," hinting at a token-efficiency-focused model with an unverified name, with no official corroboration.

Why it matters

First-hand quota perceptions from heavy paying users are a key signal for whether OpenAI quietly adjusted quotas; if true, it directly affects the practical value of premium plans like Pro/5x, warranting continued attention to further user reports and any official response.

Episode 9 · Codex Users Report Runaway Quota Burn, Blame New Astra Model (2026-09-16, 8 posts)

From September 16 to 18, OpenAI Codex users flooded channels with complaints about abnormally fast quota consumption, pointing to the new Astra model and a suspected tightening of rate limits; OpenAI has made no public response.

Confirmed

  • On September 16, Plus user crunchy-rabbit ran a controlled test: the first task of the day — asking Sol Medium to look up a single name — burned 8% of the 5-hour quota and 1% of the weekly quota, whereas a week earlier the same query could run dozens of times without hitting limits.
  • Pro user dontshootog said the 5-hour limit returned after Astra's launch, with usage running at 2.5x Sol's, losing 2% quota every 5 minutes; he judged Astra less than 10% more efficient than Sol, making it a bad deal.
  • Reddit user seasonedatheist reported the $100 Codex plan's weekly quota drained in about 3 hours on Astra, and found Anthropic's (which they called Fable) limits noticeably better.
  • Developer altryne measured over 80% of the weekly quota consumed in a single day after reset, using medium non-fast mode, questioning the "for everyone" marketing.
  • Hugging Face engineer NielsRogge publicly complained Codex had become unusable — generating just two paragraphs triggered the usage limit again — and tagged Codex lead Thib.
  • Developer ElitzaVasileva (relayed by Jason Kneen) reported 7% quota left with 3 days to reset, no spare resets, and the 20x tier paused for purchase amid a large code migration; Reddit user Junra described the same squeeze: consumption spiked after reset, an added 5x pack drained at 5-10x weekly speed, and even buying another 20x was blocked by OpenAI.
  • Another user joked that the $200 subscription's 7-day Astra allowance could fit in a small bag.

Unconfirmed

  • The exact cause of faster burn (billing coefficient changes, tighter quotas, or Astra's inherent consumption) remains user speculation; OpenAI has not explained.

Why it matters

  • The issue spans all paying tiers from Plus to $200, with high tiers both rate-limited and blocked from purchasing more, directly hurting heavy workflows; several users are already comparing Anthropic's limits favorably, which could drive subscription churn.

Episode 10 · Users Flee $200 Codex Plan as Credits Burn in Days, Claiming Claude Offers 5x Value (2026-09-27, 3 posts)

Frustrated heavy users report the $200 Codex plan's credits now last barely two days even with careful usage, and many are switching to Claude's同等价 plan, which they feel offers about five times the usable capacity with Opus 5.5.