A journalist discovered repeated $5 OpenAI charges that added up to nearly $500 after an agent in the Codex desktop app entered a loop following a reinstall. The loop repeatedly processed data on a remotely accessed Mac Mini while the user was abroad with spotty connectivity.
What happened
The episode began on a Sunday morning when the author, returning from two weeks in Europe and jet-lagged, found OpenAI charging $5 multiple times — in some cases more than 20 times a day. The author had been using the OpenAI Codex desktop app for months; it had helped both personal and professional work and contributed to the launch of a new product called Semafor Intelligence. Prior to travel, the author’s AI spending was low and they had not set limits on Codex’s automatic $5 credit top-ups.
About a month earlier a software glitch had erased all Codex chats. After deleting and reinstalling the app, the author asked Codex to revive an ambitious project. That request somehow sent an agent into a loop that kept reprocessing a set of data, consuming tokens continuously. The process ran on a Mac Mini accessed remotely, and because of consistently spotty data connection overseas, the author was not checking the machine.
Codex later explained that the agent had interpreted the project prompt “too literally”; the prompt in question had been written by Codex itself. The resulting bill was just shy of $500 — and, the author notes, it could have been worse.
Attempts to get help
The author then asked Codex to contact OpenAI’s customer service to explain the situation in hopes of a refund. That effort devolved into the author relaying messages between Codex and OpenAI’s customer-support AI chatbot. No refund was issued, and the author did not speak to a real OpenAI employee until contacting the company’s PR department. The author describes himself as the only human actively involved in the interactions.
Takeaways and wider context
The author acknowledges the bill was avoidable: mistakes included excessive trust in an unproven technology, granting it too much control and wallet access, and failing to set spending limits. The author compares the current state of the software to a fantasy-camp version of professional tools rather than fully mature developer-grade systems.
This is not only an individual issue. At a corporate level, concerns about token costs are significant: the chief data and analytics officer for JPMorgan’s Payments division told Semafor last month that token costs are exceeding some employees’ salaries.
Practical advice from the author is simple: if you experiment with AI tools, set limits on automatic credit top-ups to avoid unexpected charges.
