OpenAI has rejected claims that it changed Codex usage limits without informing developers, with Codex lead Thibault Sottiaux pointing to third-party “sub2api” tools as the explanation after paying users reported that their weekly Codex allowance was draining far faster than the plans they bought. Codex is OpenAI’s agentic coding tool, and it draws on the weekly allowance attached to a ChatGPT subscription — a flat-rate model distinct from the metered, pay-as-you-go billing of OpenAI’s API, and the same subscription-versus-API line that “sub2api” setups cross.

Did OpenAI reduce its usage limits?

Thibault Sottiaux, who now runs both ChatGPT and Codex as OpenAI’s head of core products, according to a Wired interview, addressed complaints from developers that their weekly Codex allowance was draining quickly in a widely shared X post.

Sottiaux wrote that adjusting usage caps is not something the company does without talking to the community and being transparent about it. When his team examined the accounts that were burning through allowances, he said many of them were running “sub2api” setups, which repackage a ChatGPT subscription so it can be called like OpenAI’s metered, pay-as-you-go API. OpenAI’s own documentation, however, notes that subscription access and API-key access are billed on separate tracks.

On OpenAI’s developer forum, a Codex 20x subscriber posting as anil.c1 wrote on August 14 that he was using up his weekly allowance in an almost five-hour window, despite doing the same amount of work with the basic Codex Desktop and no extra tools. Four days later, he posted again to say he had canceled his plan.

Another user, “plutavian,” said a brand-new 20x account had a weekly limit of about $200, while an older account on the same plan still had over $2,000.

The stakes are competitive as well as financial: Codex is one of several subscription-based AI coding agents — Anthropic’s Claude Code and GitHub Copilot are others — and usage allowances are one of the yardsticks developers use to compare plans in that market.

Is the OpenAI usage problem only caused by third-party tools?

A developer named Ayaan Lashari built a free app called NerfTrack, published on GitHub, that reads Codex usage data and displays the weekly value in dollar terms — a workaround for the fact that OpenAI only shows a percentage bar.

An investigation by Kingy AI described NerfTrack’s claims as “supported but unproven,” pointing out that a smaller allowance and faster spending look exactly the same on a rounded percentage bar, making it hard to tell which one is happening. The report also cited an August 20 post by Alex Getman, which claimed a Plus allowance had dropped from about $160 to $80 in API-dollar terms.

This is the second time the issue has come up in two months, and Sottiaux has addressed usage complaints repeatedly in that period. On June 30, Sottiaux held a Sunday “warroom” meeting to explain that the first incident occurred because Codex was doing extra work behind the scenes by mistake: automated review tools and helper subagents were sometimes running twice or trying too hard to fix errors, consuming more usage than intended. He also said the dashboard showed some activity that was never charged, and confirmed that fixes were put in place and everyone’s usage limits were fully reset.

Then, on July 28, according to Kingy AI, he addressed another round of complaints, saying GPT-5.6 Sol makes more tool calls and runs longer than older models, and that OpenAI had adjusted it so that normal use lasts about 18% longer. He also denied that the company had cut any subscription usage limits.

The real problem is the confusion about how the usage limit works. OpenAI’s help center says Codex, ChatGPT Work, and related tools all share the same allowance. Usage also depends on many factors, including the model being used, where the task runs, how complex it is, as well as the context, reasoning effort, speed, and the tools used.

For example, Kingy AI noted that Fast Mode runs GPT-5.6 about 1.5 times faster but spends credits at 2.5 times the normal rate. When the issue flared in June, OpenAI’s remedy was a full reset of every user’s limits; so far this time, the public response has been the sub2api explanation, with no reset or change to the percentage-only dashboard announced.