--- title: "OpenAI Quietly Tested Charging for Codex Resets. Five Days Later It Gave One Away Free." date: 2026-08-23 tags: ["openai","codex","chatgpt-work","pricing","industry","agentic-workflows"] categories: ["Industry","AI Tools"] summary: "Around August 16, OpenAI was caught testing paid usage resets for Codex — $5-8 for Plus, up to $80 for Pro — discovered in app checkout code rather than announced. Five days later, on hitting 20M Codex users, OpenAI credited every paid user a free 'banked' reset instead — one that rolled out roughly a day late and, per user reports, drained within hours for some." --- ![OpenAI Quietly Tested Charging for Codex Resets. Five Days Later It Gave One Away Free.](/images/openai-codex-paid-resets-then-free-milestone-reset.png) Two OpenAI Codex stories landed five days apart in the past week, and read together they say more than either does alone. On August 16, a limited pay-to-reset test surfaced in ChatGPT's own checkout code — not an announcement, a discovery. On August 21, OpenAI announced Codex had crossed 20 million active users and credited every paid Codex and ChatGPT Work user a free reset as a celebration. The second story reads like generosity. Read next to the first, it reads more like a company that tried charging for the thing it was simultaneously giving away for free a few days later, and got caught testing the paid version first. ## The Paid-Reset Test Nobody Announced OpenAI didn't publish a blog post or changelog entry for this. It was found by the AI-tracker account @testingcatalog, who spotted pricing tiers embedded in ChatGPT's public checkout configuration: Plus-tier resets priced at $5-8, Pro Lite at $25-40, and full Pro at $50-80 — a wide enough range that it reads like an unfinished A/B test rather than settled pricing. Tom's Guide subsequently confirmed the low end was live for at least some Plus subscribers as an actual $8 in-app prompt, and got an OpenAI spokesperson to confirm the company's general intent — described as "exploring ways for users who hit their caps to buy additional usage" — without confirming specific figures. Worth being precise about what this is and isn't: no OpenAI blog post, no on-the-record pricing confirmation, no named OpenAI executive putting their name on it. It's a silent test, discovered via app code, with a spokesperson's vague confirmation layered on after the fact. That's a meaningfully weaker sourcing bar than this blog usually requires for a standalone claim — flagged here explicitly because the contrast with what happened five days later is still worth drawing, not because the paid-reset test itself is airtight. ## Then: 20 Million Users and a Free Reset On August 21, Codex lead Tibo Sottiaux posted on X that Codex had crossed 20 million active users — up from 15 million roughly a week earlier and 10 million on July 21, a genuinely fast growth curve by any measure. To mark it, he announced every paid Codex and ChatGPT Work user would be credited a free "banked" reset, promising it would land by 8pm PST that day. (X's own paywalling of fetcher access means this post can't be quoted verbatim here with full confidence — the wording above is corroborated across multiple independent search results and a tracker site pulling directly from his feed, treated as well-corroborated paraphrase rather than a verified direct quote.) The 8pm deadline came and went for a meaningful number of users without a reset landing. This is where the story is easy to oversimplify: it wasn't a clean no-show. OpenAI's own community forum shows a staggered rollout — some users had resets by that evening, others were still waiting, and OpenAI's own follow-up comments pointed to a same-day-but-timezone-confused distribution window rather than a hard failure. By the following day, Sottiaux confirmed on X that the reset had "landed" for everyone. Call it a roughly 18-24 hour miss against the stated deadline, not a broken promise outright — but a miss all the same, on a self-imposed deadline for a feature meant to generate goodwill. What happened after it landed undercuts the goodwill further. Multiple users on OpenAI's own developer forum reported the reset drained within hours of ordinary use — one described theirs going from full to zero within six hours of what they characterized as routine work, not abuse. Another asked the obvious question directly on the forum: what's the point of a single reset if the underlying rate limits run out just as fast as before. Business/ChatGPT Work users separately worried they'd be excluded from the milestone benefit the way they reportedly had been from prior ones. ## The sub2api Deflection Sottiaux's posts that same week also addressed a separate complaint thread — users reporting Codex usage limits behaving inconsistently — by attributing much of it to "sub2api," a category of open-source tools that let people share or proxy a single ChatGPT subscription's access as if it were API access across multiple users. That's a real phenomenon (the tools exist publicly on GitHub), and it's a plausible explanation for some of the anomalies. But forum and X users pushed back specifically on the framing, several stating they use only the official client and still saw fast quota drain — rejecting the sub2api explanation as insufficient to cover what they were seeing. ## Why the Sequence Matters More Than Either Story Alone Neither the paid-reset test nor the free milestone reset is, on its own, a major story — a silent pricing experiment discovered in app code is common industry practice, and a promotional reset that ships a day late is a minor operational miss. What makes the sequence worth writing about is the five-day gap between them: OpenAI tested monetizing the exact mechanism it turned around and gave away as a customer-goodwill gesture, using it as a lever for both purposes in the same week, on a product that just crossed 20 million users and is presumably under real pressure to convert some fraction of that growth into revenue. This is a genuinely different posture than the up-front, published-changelog cadence this blog has tracked for Claude Code, where governance and rate-limit changes — subagent caps, budget flags, backgrounding rules — land as dated CHANGELOG.md entries with stated behavior, not as discovered checkout-code experiments a tracker account has to dig up. It's a small data point, not a verdict on either company's overall pricing philosophy, but it's a legible one: one of these products tells you what changed and when; the other one you find out by reading its checkout page's source. No mainstream tech outlet picked up either half of this story as of publication — it's still living in developer forums, X, and one tracker-account discovery, which is itself worth naming plainly rather than dressing the story up as more broadly confirmed than it is. --- **Sources:** - [OpenAI is testing a pay-to-reset limits feature](https://www.tomsguide.com/ai/openai-is-testing-a-pay-to-reset-limits-feature-heres-what-it-could-mean-for-your-chatgpt-usage-quota) — Tom's Guide (paid-reset test, OpenAI spokesperson statement) - @testingcatalog on X, Aug 16, 2026 (original discovery of paid-reset pricing tiers in app checkout config) - @thsottiaux (Tibo Sottiaux) on X, Aug 21, 2026 (20M users announcement, 8pm PST reset deadline, sub2api explanation) - [OpenAI Developer Community: "20 million Codex users — a free banked reset for everyone"](https://community.openai.com/t/20-million-codex-users-a-free-banked-reset-for-everyone/1391683) — user reports on rollout timing and reset drain speed