Codex User Claims a Screen Review Request Spawned 826 Subtasks and $78,000 in Charges

A user of OpenAI’s coding tool Codex published a post on September 26 (local time). In it, the user claims that a long chain of tasks they never asked for was executed, running up roughly $78,000 in costs. The user says that on July 10, working inside the VS Code editor, they asked for a review of the layout and usability of a product screen. This is not an announcement from OpenAI. It is an allegation of harm made by a single user.

One screen review, 826 tasks

The user says they went back through the logs left on their own computer. According to that account, 826 subtasks, each with its own distinct ID number, had been generated underneath the original request. Some of them, the user claims, touched internal service features and release preparation work that had nothing to do with reviewing a screen. In the end, the issue comes down to one question: did the scope of what was actually executed match the scope of what the user asked for?

The user adds that the original task and the subtasks ran on different models, and that the recorded depth of reasoning applied to generating answers differed as well. That said, this entire analysis rests on logs stored on the user’s own machine. None of it has been confirmed against OpenAI’s server-side records. A high task count and differing settings alone are not enough to conclude that anything ran without authorization or that the product is defective.

Where does the $78,000 figure come from?

The amount the user put in the post’s title and the billing total the user later tallied up do not match. The user writes that adding up 162 paid invoices produced a figure of $79,664.88. That number needs to be kept separate from the roughly $78,000 cited in the title. The crux of verifying the costs is how each of those invoices maps to the tasks that branched off from the screen review request.

Nor can the usage figures recorded in the task logs be converted directly into billed amounts. The user acknowledges as much, saying the usage totals left on their computer cannot be translated into actual charges. They also say some detailed execution records could not be found on the machine. Still, the fact that records are not visible is not by itself grounds to conclude that someone deleted them.

The user says they contacted OpenAI support and received a response to the effect that the credits had been consumed. That is the user’s own relaying of the reply, not the findings of an official investigation explaining why the tasks were created or whether each individual charge was justified.

So what needs to be verified?

Codex is an AI agent, meaning it carries out work on its own. It can be used in a code editor, in the terminal (the screen where commands are typed), and in the cloud. A cloud version was previewed in May 2025, and the tool was formally released in October of the same year. OpenAI also maintains Korean-language product documentation. Of course, the existence of Korean documentation says nothing about the terms attached to this particular account at the time of the incident, nor about the quality of the tool’s work in Korean.

The question raised by this claim is a simple one: can users see how far a job handed to an AI will expand, and how much it will cost? Settling the facts requires lining up three things: the original request, how the subtasks came to be created, and the billing records. For Korean users too, the fees attached to their particular usage pattern and the conditions under which tasks get approved are important issues. From here, what is worth watching is OpenAI’s explanation based on server-side records, and how the costs are ultimately handled.