FULL STORY

The Codex Quota Saga: runaway usage, outages, and reset chaos

Starting mid-September, Codex subscribers reported quotas draining abnormally fast, followed by a full outage, surprise hard resets, timezone bugs, and denied grandfathered credits, fueling ongoing backlash over OpenAI's billing transparency.

2026-09-12 ~ 2026-10-05 · 11 episodes · 72 posts

Episode 1 · Users report Codex credits burning faster than usual (2026-09-12, 5 posts)

From September 12 to 14, multiple users reported that their OpenAI Codex credits were draining unusually fast, suspecting a change in usage metering or backend behavior; OpenAI has offered no official explanation.

Confirmed

  • chriscarm posted on September 12 that after his credits reset that day, he burned from 100% to 0% in under 56 minutes, calling it his personal record and illustrating real consumption under heavy coding.
  • A Reddit user reported that starting about a week earlier, routine tasks on the Codex 5x plan (model 5.6, not heavy coding) drained roughly 1% every 10 minutes, with the reset cycle apparently changing to 10 days; adding coding tasks (model 6.0) consumed heavily on day one.
  • Another user said that just one day after a reset, a single /goal run (Astra high) consumed 16% of credits, asking the community whether others saw the same.
  • TheMoonMidas asked on September 14 how many people noticed Codex credits draining faster than the previous week, sparking speculation about metering changes.

Unconfirmed

  • Whether faster consumption stems from metering changes, backend behavior, or heavy usage remains user-side reports with no official OpenAI response.
  • Some data points (e.g., the 10-day reset cycle) come from a single user without cross-validation.

Why it matters

  • If the metering basis changed, it directly affects actual usable quota and perceived cost for Codex subscribers, especially those on the 5x plan.
  • chriscarm's 56-minute record plus multiple reports highlight quota pressure in high-intensity coding scenarios and heavy users' sensitivity to quota policy.

Episode 2 · Paying users complain Codex subscription limits drain far too fast (2026-09-18, 7 posts)

Several paying users on Reddit are collectively complaining that OpenAI Codex subscription limits drain far faster than expected — even the top 20x tier barely supports heavy use — and are urging OpenAI to publish clear usage quotas.

Confirmed

  • @NoTwist678 tested the 20x tier: one hour of usage consumed 5% of the weekly limit, implying heavy use cannot last a full week; he mocked OpenAI for running "beast mode" on its own users.
  • @craigbalding said his 20x quota was wiped out in one go by astra xhigh and could only be reused for Codex after the Tuesday reset; a new 5x subscription also exhausted its weekly limit within two days. He shared multi-model cost-saving mixes, such as offloading work to terra medium.
  • Relayed by @NYCounihan, user AAAzzam reported running out of quota with 4 days left in the cycle and no reset, the first time in months, and complained the Astra tier is expensive.
  • @AIandDesign reported that even in Sol mode, the $200 Codex plan burns through quota quickly.
  • Relayed by @CtrlAltDwayne, DanielLockyer complained Astra is essentially unusable on a 20x subscription, let alone lower tiers, and urged OpenAI to follow Anthropic's approach with Fable by giving subscribers clear usage quotas.
  • @umeshonai calculated that six 5-hour sessions would exhaust the weekly quota, arguing Claude offers better value for heavy coding users.
  • @StewartalsopIII noted that when subscribing to both Codex and Claude, the two reset cycles are almost fully misaligned yet consumed together: using Astra can drain Codex quota on day one with a 6-day wait, leaving Claude quota for Fable only in the last two days, with limited marginal benefit from premium models.

Why it matters

  • The episode suggests OpenAI's quota calculation is seriously mismatched with real-world consumption by heavy coding users, hurting Codex's viability for Pro-tier users.
  • Users repeatedly compare Anthropic/Claude's quota transparency and pricing; without adjustments, OpenAI risk losing share among heavy coding users.

Episode 3 · Pro Users Swarm With Complaints Over Codex and Astra Usage Limits (2026-09-19, 14 posts)

From Sept 19–21, high-tier OpenAI subscribers took to X and Reddit to complain en masse about Codex and Astra usage limits: quotas allegedly inconsistent with marketing, draining unusually fast, and resetting unpredictably, with some accounts suspended mid-quota. Some users have cancelled or sought EU refunds; OpenAI has not officially responded.

Confirmed

  • Developer adrianhqs tweeted that "Codex users are being scammed," attaching screenshots suggesting actual usage/quota mismatch with marketing (09-19, m1).
  • Developer @buildwithrajath attacked the reset mechanism: hitting limits mid-task, watching usage countdowns, killing unfinished agents, and guessing when quota returns (m2); another user criticized this "reset culture" for making paying users change behavior out of fear of wasting quota, calling for predictable credits (m3).
  • $200/month Pro user btsouth said quota hit 97% within 24 hours of reset with 6+ days to wait, cancelled the subscription, and now plans work around reset windows (m5).
  • User AIandDesign reported generous early quotas that lured him into a $200 bundle, only to drop to 67% remaining within 48 hours, pushing most real coding to Meta Muse (m7); he separately said that with barely any use he burned 20% of the $200 plan in a day, questioning whether the $100 tier affords only 20 minutes and tagging Codex lead Thibault (m6).
  • A $100/month Pro user reported persistent model downgrades and tightened Codex token limits, with available usage feeling "essentially random" (m4); Pro user alexcovoeth reported reset quota draining abnormally fast (79% toward the next reset a day after using one) and suspected quality degradation in Astra 6 with more errors on multi-task generation (m12).
  • An EU Reddit user said Codex quota was cut below the $20 tier after paying over €100 for Pro, and asked how to seek an EU refund (m8).
  • Researcher kfountou said his account is periodically suspended with 90% quota remaining, suspecting rate-limiting for excessive compute use (m9); RL scholar Csaba Szepesvári exhausted his entire token quota while preparing a graduate course, after earlier being told his chosen model was at capacity (m10).
  • A Reddit user complained of being "stuck on the third 5-hour window," handing remaining assembly and verification work to another session, with resonant comments (m11).
  • User iamrobotbear reported Astra quota draining noticeably faster over three days, hitting 100% on the 20x plan in two days, suspecting silent tightening (m14); developer melvindvivas called Codex "unsustainable" under current limits (m13).

Unconfirmed

  • All accounts are one-sided user claims; "silent quota cuts / stealth downgrades / compute-based rate limiting" have no official confirmation from OpenAI, which has not responded.
  • Faster consumption may also reflect model tier choice (e.g., mid-tier Astra), workload, or normal tuning—no controlled data exists.

Why it matters

  • High-tier Pro/Max subscriptions anchor OpenAI's coding revenue; widespread "paid but afraid to use it" experiences directly undermine the subscription value proposition.
  • Users are already voting with their feet—cancelling, seeking EU refunds, or switching to competitors like Meta Muse; continued quota opacity could accelerate churn.
  • The episode touches marketing-vs-reality and stealth downgrade concerns that could trigger further disputes under EU consumer protection rules.

Episode 4 · Users Accuse OpenAI of Repackaging Terra as Sol While Cutting Usage Limits (2026-09-23, 2 posts)

Developers accuse OpenAI of repackaging Terra as Sol with cheaper output prices but worse usage caps, while Codex limits have shrunk so much that some burn a week's quota in two days.

Episode 5 · Users Slam OpenAI Codex for Runaway Quota Consumption (2026-09-24, 4 posts)

Multiple users reported runaway quota consumption in OpenAI Codex, with one claiming nightly tasks burned 30% of a $200 plan and another saying its token efficiency is a tenth of Claude Code's, leaving Pro limits spent in a day.

Episode 6 · OpenAI Codex suffers global outage, restores service and resets paid usage (2026-09-26, 28 posts)

On the evening of Sept 25 through Sept 26, OpenAI Codex suffered a full outage affecting Web, API, CLI and the VS Code extension, with large numbers of users receiving 401 Unauthorized: Incorrect API key errors. The incident was confirmed at 10:58 and recovered about an hour later; Codex lead Thibault Sottiaux announced a usage reset for all paid Codex and ChatGPT users and apologized. No official root cause has been given, which is why the incident remains notable.

Confirmed

  • The OpenAI status page showed the outage affecting Codex Web, API, CLI and VS Code extension, confirmed from 10:58 and recovering around 11:47; during the incident users were advised to log in with an API key, which was verified as a workaround (reachvb). The status page later declared full operational status with a reset record, publishing availability data: APIs 11 components at 99.96%, ChatGPT 15 components at 99.66%. The next day dfinke and others confirmed API credits had been reset.
  • Many users (iamgarffi, aidenybai, Jeremy Philippe, TheGrog, JustAnotherUser80 and others) reported 401 invalid-key errors across Codex; a ChatGPT Pro user hit repeated 401s on Codex Desktop for Windows even when signing in with ChatGPT, with macOS also affected. Others saw "workspace routing discovery timeout" errors. One user had 4 long-running tasks all stuck in infinite "Reconnecting"; another was mid-refactor and forced to stop.
  • Errors referenced an sk-svcac-prefixed key users never set, leading some to suspect an OpenAI-side authentication failure.
  • The status page initially showed no incident when iamgarffi posted; the official OpenAI Devs account then confirmed the outage, said fixes were underway, and apologized after resolution without disclosing cause or scope.
  • After recovery, Thibault Sottiaux announced a usage reset for all paid Codex and ChatGPT users; during the downtime many users flooded his replies demanding resets (rudrank). He joked the team keeps a "backup Codex" for outages.
  • The incident became meme material: Yuchen Jin's "all of Silicon Valley" reaction, engineer aidanclark initially blaming his airplane WiFi, clairevo's "pov: codex is down", and theo not even noticing; some heavy users argued for local fallback models for critical systems so cloud coding outages don't halt work.

Unconfirmed

  • The exact root cause (authentication-side misconfiguration or otherwise) remains officially unexplained; the sk-svcac key observation is user-side.

Why it matters

  • Codex is a key OpenAI developer tool, and authentication failures halt entire workflows at scale; the incident also exposed status-page lag behind real outages, relevant for operators and dependents.

8 more related posts →

Episode 7 · Codex Outage Burns Through Paid Quotas, No Compensation (2026-09-26, 2 posts)

After an roughly 8-hour outage, OpenAI's Codex restored service but did not compensate users' usage quotas, with some burning 6% of their weekly GPT-6 Pro allowance just on troubleshooting, sparking billing complaints.

Episode 8 · OpenAI Codex Users Protest Sudden Quota Hard Resets (2026-09-27, 2 posts)

OpenAI Codex users are protesting unannounced 'hard resets' that wiped remaining quota—some lost 60% of their usage and must wait another 7 days—while others found reset times silently moved earlier, defeating plans to spend remaining quota.

Episode 9 · User Math: Codex Reset 15 Times in 60 Days, 174% Extra Quota (2026-09-30, 2 posts)

A Reddit user tallied Codex quota grants over 60 days, finding 11 full resets plus 4 banked resets—about 174% more quota than a normal weekly allocation would provide.

Episode 10 · OpenAI Users Fury Over Botched Codex Quota Resets (2026-10-03, 3 posts)

OpenAI's quota resets have angered users: Pro weekly limits were wiped early, Business accounts were skipped in the global Codex reset, and a $500 subscriber saw usage zeroed overnight, sparking complaints about unpredictable metering.

Episode 11 · Codex Quota Resets Ignore Time Zones, Users Lose Credits (2026-10-05, 3 posts)

OpenAI researcher Elie Bakouch and other users found that Codex quota reset times are not adjusted to users' time zones, causing credits to expire unexpectedly before the displayed deadline.