OpenAI’s Codex agent is under fire from users who say the OpenAI Codex logging is chewing through SSD write cycles far faster than expected. The complaint is simple: the service logs so aggressively that it may shorten the life of a drive, turning a debugging feature into a silent hardware tax.
One developer who relies heavily on Codex said the service wrote about 37 TB of data in 21 days. If that pace holds, it works out to roughly 640 TB a year – enough to burn through the rated write endurance of some consumer NVMe drives in less than a year. That is awkward timing for a category where endurance is one of the few specs that still separates the sensible from the shiny.
How much SSD wear Codex may cause
Users have been doing the math themselves. A commonly cited formula for the damage is: terabytes written multiplied by the SSD price divided by the drive’s rated endurance in terabytes. Using an average $200 SSD, those extra 37 TB of writes would cost about $12.33, though Codex reportedly calculated $38.64 for one affected user based on the price of a 2 TB Samsung 990 NVMe.
- 37 TB written in 21 days
- About 640 TB written per year at the same pace
- Samsung 9100 PRO 1 TB rated for 600 TBW
- Estimated cost of 37 TB of extra writes: $12.33
OpenAI says it is fixing the logging
OpenAI has acknowledged the issue and says it is working on a fix. The company’s explanation is that Codex writes heavily so engineers can diagnose problems more easily, but the system appears to have become more chatty on disk than intended. That trade-off is familiar in software, yet it gets expensive fast when the bill is paid in NAND wear instead of cloud storage.
Users have been reporting the problem through GitHub for months, and OpenAI has already made some changes without eliminating it. If the complaints are right, the bigger headache is not a single angry developer but the cumulative effect across a large user base – the kind of quiet cost that can run into millions before anyone outside the affected group notices.
Why the SSD backlash is spreading
This is also a reminder that AI agents do not just consume compute; they can generate real wear on everyday hardware if logging gets sloppy. Competing products are increasingly being judged not only on output quality but on how much hidden overhead they create for the people running them. The next test for Codex is straightforward: cut the writes, or keep explaining why a debugging tool is behaving like a stress test.