OpenAI Codex Wears Out SSDs with Massive Writes: 640 TB/Year

What Happened?

A bug in the logging implementation of OpenAI's Codex coding agent is causing a massive volume of writes to users' solid-state drives (SSDs). According to a GitHub report identified as #28224, developer Rui Fan, a member of the Apache Flink project management committee, detected that his SSD had accumulated 37 TB of writes in just 21 days of continuous Codex use. This extrapolates to approximately 640 TB per year, equivalent to 640 full write cycles on a 1 TB SSD. To put it in context, a typical consumer SSD has a endurance rating between 300 and 600 TBW (terabytes written), so this write rate could exhaust the drive's guaranteed lifespan in less than a year. The problem is not new in the industry: in 2014, a similar bug in Google Chrome's file system caused excessive writes to users' SSDs, though the impact was smaller (around 10 TB per year). The difference here is the magnitude: 640 TB per year is an order of magnitude higher.

Impact on SSD Lifespan

SSDs have a limited lifespan, measured in terabytes written (TBW). For example, the Samsung 9100 PRO 1 TB has an endurance of 600 TBW. At the current rate, a Codex user could exhaust that warranty in less than a year. Another developer reported that, according to Codex's own analysis, the bug cost him $38.64 in value on his Samsung 990 2 TB SSD. The total cost to the user community from March to June 2025 is estimated to reach several million dollars, assuming a cost of $0.13 per TB written. However, this calculation is conservative: using a price of $0.33 per TB (based on a 1 TB SSD at $200), Fan's cost would be $12.33 for 37 TB. For a 2 TB SSD with higher endurance (1200 TBW), the cost per TB drops to $0.25. The economic impact varies by SSD model, but even in the best case, accelerated wear reduces resale value and shortens hardware lifespan. This is especially relevant for developers using high-end SSDs (like Samsung 990 Pro or 9100 Pro), which can cost over $300. Additionally, premature wear can cause unexpected failures, data loss, and replacement costs.

Origin of the Problem

The problem originated in February 2025, when a change was introduced to log SQLite database entries at the TRACE level, which is much more detailed than the ERROR level. Although these logs are intended to help OpenAI engineers diagnose issues, the high data volume and storage method generate much higher disk activity than expected. Ironically, Codex itself (presumably running GPT-5.3) reviewed the commits that introduced the bug, raising questions about the quality of automated review. This incident recalls other cases where AI tools overlooked critical errors, such as the GitHub Copilot bug in 2023 that generated insecure code. In that case, human review caught the problem, but here automation failed. OpenAI should implement better performance testing and disk write monitoring before deploying changes. Moreover, excessive logging is a known software issue: in 2020, a bug in Windows 10's logging system caused massive writes to SSDs, though it was quickly fixed. The difference is that Codex runs locally and the user bears the cost of wear.

Consequences and OpenAI's Response

OpenAI has confirmed that its engineers are working on a fix, and several pull requests have already been made to mitigate the problem. However, users continue to report issues. To date, no official patch has been released, though one is expected in the coming weeks. Meanwhile, users can limit Codex usage or configure the logging level to reduce writes. This incident underscores the importance of optimizing logging in AI applications, especially when running locally on user hardware. It also highlights the risks of relying on AI tools to review their own code, as they can miss serious performance bugs. For OpenAI, this bug represents a reputational risk, especially after announcing plans in December 2024 to add telemetry by default in Codex CLI (except where prohibited by law). Users now distrust that telemetry could exacerbate SSD wear. At the market level, this incident could slow Codex adoption among developers concerned about hardware durability, benefiting competitors like GitHub Copilot or Amazon CodeWhisperer, which have not reported similar issues.

What Should Readers Know?

If you are a Codex user, check the amount of data written to your SSD using tools like CrystalDiskInfo or smartctl. Consider limiting Codex usage until an update is released. For future SSD purchases, prioritize models with high TBW endurance (e.g., 1200 TBW or more). This case also serves as a lesson for developers: excessive logging can have a real cost, both in hardware wear and economic value. Additionally, it is advisable to regularly monitor SSD health and perform frequent backups. OpenAI should provide a tool to estimate the wear caused by Codex and offer symbolic compensation, such as platform credits. In the long term, the industry needs standards for efficient logging in local AI applications, similar to the Open Source Security Foundation (OpenSSF) guidelines for secure software. Until then, users must be proactive in protecting their hardware investment.