Three days ago, this blog covered OpenAI getting caught testing paid Codex usage resets, then giving one away for free five days later when Codex crossed 20 million users. That piece closed on an open question: which direction was OpenAI actually heading on usage limits — toward monetizing scarcity, or toward genuine generosity? On August 25, the answer arrived, and it’s neither. OpenAI just put the scarcity back.
The full arc, dated#
Lay out the primary-sourced timeline and the pattern is hard to miss:
- July 12, 2026: Codex lead Thibault “Tibo” Sottiaux posted on X that OpenAI was temporarily removing the rolling 5-hour usage limit on Codex and ChatGPT Work for Plus, Pro, and Business plans — framed as one of three simultaneous changes, alongside a mid-cycle usage reset for everyone and efficiency gains from GPT-5.6 Sol reducing quota burn per task. The weekly cap stayed in place as the real ceiling; removing the 5-hour window meant a bigger available runway within that week, not unlimited usage. No end date was published — “temporary” with nothing more specific.
- Around August 16: A limited pay-to-reset test surfaced in ChatGPT’s own checkout code, discovered rather than announced — Plus-tier resets priced at $5-8, Pro tiers up to $80, per app-tracker @testingcatalog’s findings and Tom’s Guide’s follow-up confirmation with an OpenAI spokesperson.
- August 21: Codex crossed 20 million active users. Sottiaux announced every paid Codex and ChatGPT Work user would get a free “banked” reset to mark it — a rollout that landed roughly a day late for a meaningful share of users and, per multiple forum reports, drained within hours for some of them.
- August 25: Sottiaux confirmed on X that the rolling 5-hour limit is back for ChatGPT Plus subscribers, across both Codex and ChatGPT Work, on top of the existing weekly allowance. His stated rationale: smoothing compute load while keeping the weekly limit “generous,” plus a pointed detail — casual users were “accidentally eat[ing] through their whole week’s usage” without the 5-hour guardrail, then showing up confused. Pro subscribers at both the $100 and $200 tiers stay exempt “for the upcoming months.”
Six weeks, almost to the day, from “temporarily removed, no end date” to “reinstated in full.” Whatever “temporary” meant on July 12, it turned out to mean a specific, if unannounced, number of weeks — and it took a secretly-tested paid-reset scheme and a PR-timed free giveaway to fill the gap in between.
What actually changed, and what didn’t#
Be precise about the mechanics, because it’s easy to overstate this as OpenAI “cutting” something it never fully gave. Plus users always had a weekly cap; the July 12 change only removed the 5-hour rolling sub-limit within that week, giving heavy same-day users a bigger short-term runway. August 25 restores that sub-limit — Plus users hitting either the 5-hour window or the weekly cap now have to wait for a reset or buy additional credits, the same two-limit structure that existed before July 12. Pro users ($100/$200 tiers) are untouched by either the July change or the August reversal; they were uncapped on the 5-hour window before, during, and after this whole arc, and Sottiaux’s own wording — “for the upcoming months” — is the only signal about how long that holds.
So the net effect for a Plus subscriber over six weeks: a temporary widening of the same-day allowance, a brief window where OpenAI tested charging for more of it, a one-time free top-up timed to a growth milestone, and then a return to where things started. Read uncharitably, that’s six weeks of goodwill theater bookending a company’s own admission that the un-metered version costs more compute than it wants to keep giving away below the Pro tier’s price point. Read charitably, it’s a company that ran an actual experiment, measured what casual users did with more headroom (burned through their whole week and got confused, per Sottiaux’s own account), and reverted based on that signal rather than ideology. Both readings are consistent with the same facts, and OpenAI’s own explanation supports the charitable one — but only the uncharitable one explains why the reversal announcement came from an X post rather than a blog post with the same production values as the original removal announcement.
The contrast worth drawing#
This blog covers Claude Code’s rate-limit history too, and Anthropic isn’t a company that has never touched a usage limit — it doubled Pro/Max/Team rate limits in May off the back of new compute deals and has adjusted pricing tiers more than once this year. The difference in this specific pattern isn’t “one company touches limits and the other doesn’t.” It’s that OpenAI’s version compressed a full cycle — generous change, quietly tested monetization of the thing it had just made generous, a goodwill correction, then a reversal — into six weeks on a single product surface, announced entirely through one executive’s X account rather than through OpenAI’s own blog or changelog at any step. Compare that to how this blog covered Claude Code’s Auto mode default rollout landing exactly as scheduled with a dedicated Anthropic blog post and a published dataset behind the decision, or today’s Auto mode follow-up release shipping concrete tooling in response to concrete user complaints, verifiable against a public changelog at every step.
If you’re planning team workflows or CI budgets around either product, the practical takeaway is the same one this blog keeps returning to: verify usage-limit claims against the vendor’s own changelog or blog, not a single executive’s social post, and build in slack for the fact that “temporary” generosity from a compute-constrained vendor has a real chance of being exactly that — temporary.
Sources: 9to5Mac, “OpenAI restores 5-hour Codex and Work limits for ChatGPT Plus users” (secondary, citing Thibault Sottiaux’s X posts directly); this blog’s own prior coverage of the July 12 uncapped change and August paid-reset/free-reset saga and Claude Code’s Auto mode default rollout for context.