Codex Hits 20 Million Users While Quotas Mysteriously Drain

On August 21, 2026, OpenAI’s engineering lead for Codex, Thibault “Tibo” Sottiaux, announced that Codex and ChatGPT Work had crossed 20 million active users. To mark the milestone, OpenAI credited every Codex and ChatGPT Work subscriber with a banked usage reset they can redeem whenever they like. It should have been a straightforward victory lap. Instead, the celebration landed directly on top of one of the most persistent headaches in the product’s short history: developers across X, the OpenAI community forum, and Discord were reporting that their usage limits were draining at bizarre, seemingly impossible rates — and they were loudly not satisfied with the company’s explanation.

The growth number is genuinely staggering

Whatever else is true, Codex’s adoption curve is remarkable. The product sat at roughly 600,000 weekly active users at the start of 2026. It hit 5 million in early June, 8 million on July 14, 10 million on July 21, and 15 million on August 13 — before crossing 20 million on August 21. That is roughly a 33x increase in eight months, driven by the GPT-5.6 launch and OpenAI’s decision to fold Codex directly into ChatGPT as a unified workspace. Sottiaux, who has become the public face of Codex through a string of milestone posts on X, teased additional announcements to come.

But the pace of that growth is also the backdrop that makes the current controversy intelligible. OpenAI has been aggressively subsidizing usage to win developers: a $100 Pro tier with 5x Codex usage was introduced explicitly to undercut Anthropic’s Claude Code, and milestone resets have been handed out at nearly every growth marker — Sam Altman’s original pledge to reset limits for every million new users up to 10 million has since been extended well past its cap. When usage grows 33x in eight months while the underlying compute scales more slowly, something eventually has to give. The context window reductions that followed the 8-million-user milestone were an early signal. The current wave of limit complaints may be another.

What developers are actually seeing

The reports are specific, documented, and in some cases quantified down to the token:

• A developer paying $200/month for the Pro 20x plan reported burning through the entire weekly quota in under two days.
• A forum user described purchasing 2,500 additional credits after hitting their limit, only to watch all of them vanish within two hours — while the analytics page showed zero usage events recorded.
• A longtime subscriber said their effective weekly allowance had dropped to roughly one-tenth of what it was a month earlier: limits now hit around 38 million tokens, versus the 2–3 billion they could previously consume without issue.

The most forensically detailed account so far is GitHub issue #40067, filed August 22. A ChatGPT Plus user on Windows reported that their weekly Codex allowance — showing ~99% remaining on the morning of August 21 — hit 0% within a few hours, after which every request returned HTTP 429 usage_limit_reached. The user was running a local agent over ChatGPT OAuth, not an API key, doing routine agent tasks: scheduled briefings, file reads, integration testing. Their local logs for the period reconcile to 107 successful model calls and ~9.39 million gross input tokens (about 88.9% served from cache). Codex’s own server-side profile analytics, however, attributed 81.7 million tokens to that same day — nearly nine times what the user could verify locally. On a historically comparable day (August 12), ~96.7 million tokens had consumed only about 60% of the allowance.

The issue author asks the questions many users are now asking: is usage being double-counted, misattributed across OAuth surfaces, inflated by invisible background tasks, or has the GPT-5.6 weekly weighting simply changed without documentation? Notably, the visible 5-hour meter that once accompanied the weekly limit no longer appears on the account since the GPT-5.6 usage changes — and one theory circulating among developers is that weekly limits were accidentally applied at the old 5-hour window’s rate. That theory is unconfirmed, but it gained traction precisely because so many users observed the same sudden cliff at the same time.

OpenAI’s explanation: “sub2api”

Hours before the 20-million post, Sottiaux published a separate thread on the limit behavior. He said OpenAI does not change limits without engaging the community, that cache hit rates had in fact been worse than the prior stable state for some users that week (with an update promised), and that the team had found a pattern among many affected accounts: use of “sub2api” — a workaround that converts a ChatGPT subscription into API-style traffic that can be redistributed or shared across multiple users. He called it unsupported and said it triggers OpenAI’s fraud-prevention systems, while clarifying that official clients and open-source tools like Pi and OpenCode that support Sign in With ChatGPT are “completely fine.”

The reaction was sharp. Some users felt the explanation amounted to a blanket fraud accusation — one reply called it “unconscionable.” Many others simply reported that they had never touched sub2api, access Codex exclusively through official clients, and still see aggressive drain. The daily.dev summary of the incident put the community’s core objection plainly: OpenAI is preparing paid usage resets while limits “mysteriously” drain faster, and the sub2api account doesn’t cover the documented cases.

A pattern, not an incident

This is at least the third time in 2026 that Codex usage complaints have forced an OpenAI response. In March, OpenAI’s own status page acknowledged a bug consuming usage faster than expected. In late June, the company confirmed its abuse- and fraud-prevention systems had been incorrectly rate-limiting legitimate accounts, prompting an emergency response and a universal reset; hundreds of developers reported abnormal consumption, with some $200-plan accounts exhausting quota plus paid top-ups within hours (as covered by Business Insider at the time). And as far back as May, the team had root-caused a rolled-back optimization that damaged cache hit rates and reset limits for all accounts.

That history is why the banked reset, however welcome, is unlikely to settle anything. Sottiaux’s own 20-million post concedes the tension: the team “has not detected anything abnormal in the overall usage system,” but the investigation is ongoing.

Why this matters beyond Codex

The episode is a case study in the economics of agentic AI products. Coding agents are the first AI category where heavy daily users consume millions of tokens per session, where caching is a first-order cost lever, and where quota accounting is effectively a billing system that must be transparent and auditable. When the meter and the receipts disagree — 9.39 million locally logged tokens versus 81.7 million server-side — trust erodes fast, because developers are uniquely positioned to check the math.

It also illustrates a structural risk for every lab racing to subsidize agent adoption: usage that compounds 33x in eight months eventually collides with infrastructure that doesn’t. Resets and milestone credits paper over individual incidents, but the underlying tension — between flat-rate subscriptions and near-unbounded agent workloads — is not going away. Anthropic, Google, and xAI all face versions of the same arithmetic as they push subscription-priced agents into daily engineering workflows.

For now, the scoreboard reads: 20 million users, one banked reset per subscriber, an open GitHub issue with a ninefold token discrepancy, and a community that has heard “it’s fixed” before. The next concrete signal to watch is the promised cache-hit-rate update and whether OpenAI publishes an auditable per-surface usage breakdown — the single change that would let developers verify, rather than trust, that their quota math adds up.