OpenAI blames 'sub2api' for shrinking Codex limits, and shows no meter to check it against

OpenAI's head of core products, Thibault Sottiaux, who now runs both ChatGPT and Codex, said on August 21 that the company does not adjust usage caps without talking to the community, and pointed instead at third-party "sub2api" tools that repackage a ChatGPT subscription so it can be called like OpenAI's metered pay-as-you-go API ^[3][18][19]. The explanation is doing little work with customers, because OpenAI shows a rounded percentage bar and no clear map of what actually consumes a weekly allowance, so no paying team can audit the claim from the outside ^[4][5].

The complaints are specific. On OpenAI's developer forum, a Codex 20x subscriber posting as anil.c1 wrote on August 14 that he was burning a full weekly allowance in roughly five hours while doing the same work as before with plain Codex Desktop and no extra tools; four days later he posted again to say he had cancelled ^[20][1]. Another user, "plutavian," reported that a brand new 20x account carried a weekly limit of about $200 in API-dollar terms while an older account on the same plan still showed over $2,000 ^[21], a gap of at least tenfold on identical paid tiers ^[24]. A developer named Ayaan Lashari published a free tool, NerfTrack, on GitHub that reads Codex usage data and translates the percentage bar into dollars ^[6].

An investigation by Kingy AI rated the claim that limits were secretly cut as "supported but unproven," noting that a smaller allowance and faster consumption of an unchanged allowance look identical on a rounded percentage bar ^[7][8]. That is the whole problem in one sentence. Kingy AI also flagged an August 20 post by Alex Getman claiming a Plus allowance had fallen from about $160 to about $80 in API-dollar terms ^[22], a halving if accurate ^[25].

The sub2api story also sits awkwardly against OpenAI's own documentation, which, per the Cryptopolitan account, states that subscription access and API-key access are billed on separate tracks ^[9]. And this is the second flare-up in two months ^[10]. On June 30 Sottiaux held a Sunday "warroom" and explained the first round as Codex doing unintended extra work, with automated review tools and helper subagents sometimes running twice or over-trying error fixes ^[11][12]. He said the dashboard had displayed activity that was never charged, that fixes shipped, and that limits were fully reset ^[13]. On July 28, according to Kingy AI, he addressed another round by saying GPT-5.6 Sol makes more tool calls and runs longer, and that OpenAI had tuned it so normal use lasts about 18% longer, while again denying any cut to subscription limits ^[14].

Underneath all three episodes is an accounting surface nobody can reason about. OpenAI's help center says Codex, ChatGPT Work and related tools draw on one shared allowance, and that consumption depends on model, where the task runs, complexity, context, reasoning effort, speed and tools used ^[15][16]. Kingy AI notes Fast Mode runs GPT-5.6 about 1.5 times faster but spends credits at 2.5 times the normal rate ^[23], meaning roughly 1.67 times the credits for the same work, paid for in wall-clock time ^[26]. A team cannot forecast a $2,400-a-year seat against that ^[17].

The operator lesson is not about model quality. It is that a spend cap you cannot read is an unpriced input, and unpriced inputs get removed from build plans. Watch whether OpenAI ships a per-task usage ledger in dollars rather than a percentage bar, whether it publishes multipliers per model and mode, and whether it names sub2api enforcement as policy instead of as an explanation after the fact.