A wave of Codex users recently noticed their usage quotas draining far faster than expected. Tibo, the OpenAI executive who leads Codex, addressed the complaints directly: usage limits are not adjusted without community communication and transparency. After looking into the pattern, though, a significant share of the affected accounts turned out to be running sub2api.
He didn't soften the message: converting subscription quota into API traffic and then reselling or sharing it across multiple users is an unsupported use case that gets flagged by the anti-fraud system. Anyone using their subscription normally through Sign in With ChatGPT is unaffected — that covers the official client as well as open-source clients like Pi and OpenCode that support this login method.
How Seat-Sharing Became a Risk Signal
sub2api is an open-source relay project that funnels subscriptions to Claude, OpenAI, Gemini and Grok into a single gateway, which then issues API keys to multiple users. The repo's own description says it "supports seat-sharing to split costs more efficiently." It currently has 38,000 stars, 7,900 forks, and more than six thousand commits on its main branch.
That scale alone explains why OpenAI would run into it during troubleshooting. The disclaimer at the top of the repo is even more direct: using the project may violate upstream providers' terms of service, the developers have never authorized any form of commercial operation, and users bear the risk themselves. Whoever wrote that knew exactly what kind of territory this is.
Detecting this kind of traffic isn't hard for a risk model. A per-seat subscription account normally shows single-user, single-device behavior with an uneven rhythm tied to someone's daily schedule. Once it's routed through a gateway, the same account starts showing concurrent requests from multiple IPs, flat request density, context that doesn't carry across sessions, and load that never dips late at night. Stack those signals together and they look a lot like bulk API usage — and nothing like a person typing in an editor.
Twenty Dollars for a Two-Hundred-Dollar Plan
The economics of seat-sharing don't need a calculator: split a $200-tier subscription ten ways, and each person pays $20 for close to the full quota. Developers in China know this playbook well enough that it barely needs explaining — paywalls, exchange rates, payment-channel friction, unused quota going to waste: any one of those reasons is enough on its own.
For the vendor, the math runs the other way. Per-seat pricing assumes a natural ceiling on how much one person can use — nobody writes that much code in a day. Seat-sharing blows past that ceiling, so a single subscription's actual consumption can approach the sum of ten people's usage, and the assumption behind the pricing model collapses. That's also why the industry's countermeasure almost always starts with detecting an account's usage shape rather than raising prices — a price hike punishes honest users, while risk controls target only the arbitrage.
This isn't the first move of its kind this year. Another frontier-model vendor started requiring identity verification on flagged accounts back in July, running on the same logic: tying an anonymous quota pool back to a traceable individual.
The Loophole OpenAI Left Open
Tibo's statement carried one important qualifier: accounts using Sign in With ChatGPT are unaffected. That draws a fairly clear line around what the company actually endorses — a subscription can be used inside third-party clients, as long as the login goes through official OAuth and requests carry a real user identity, rather than being stripped down into a bare key and passed along.
For the open-source client ecosystem, that line matters. Tools like OpenCode aren't the problem — the key-relay layer is. Separating the two avoids collateral damage to a legitimate slice of the third-party ecosystem, and gives seat-sharing users a way to walk back on their own.
Codex's quota mechanics have had a rough year already. In late June, OpenAI acknowledged a backend fault where compute usage exceeded its intended baseline, draining some users' quotas abnormally, and issued a full reset afterward. Tibo announced resets often enough that the community started calling him the "cyber godfather." Frequent resets and limit adjustments have left users without a stable sense of what to expect, so any fluctuation gets read first as a quiet cutback. Pointing the finger at seat-sharing this time is, to some extent, an attempt to redirect that suspicion.
What Happens Next
So far, OpenAI has only said flagged accounts get marked — not whether that leads to throttling, mandatory verification, or a ban. Based on current reports, quota drops and forced phone verification are already happening; how common permanent bans will become is still unclear.
For individual users, the actionable takeaway is simple: switch back to official OAuth login and stop handing subscription keys to a relay layer. For anyone running a relay service, a disclaimer doesn't block a vendor's technical countermeasures, and the risk-control net will only keep tightening. This business was built from day one on the gray area of someone else's terms of service — and once that gray area gets spelled out in plain language, the ground it stood on disappears.
Sources: public remarks from OpenAI's Codex team lead, the sub2api open-source repository's description and disclaimer, CocoLoop, developer community feedback; star/fork counts and the quota anomaly and reset timeline verified against public information.