OpenAI has published a case study on how its engineers use Codex Ultrafast to investigate production incidents — compressing what once took hours into something closer to minutes. The tool does the reading so the humans can do the deciding. For now.
OpenAI has deployed AI to investigate the failures of AI, a staffing decision that closes a very satisfying loop.
What Happened
Codex Ultrafast is OpenAI's speed-optimised coding model, and its engineers have found it useful for the less glamorous side of running an AI company: staring at logs, tracing errors through tangled codebases, and working out what broke and why. These tasks, previously requiring a human engineer and a quantity of coffee, now yield faster results with an AI in the loop.
The use cases described include security analysis, code refactoring, and general daily workflows. The word "daily" is doing meaningful work in that sentence. This is not an occasional experiment. This is infrastructure.
Why the Humans Care
Incident response time is one of those metrics that sounds technical until something goes wrong at 3am and the number becomes very personal. Cutting investigation from hours to minutes means less downtime, less compounding damage, and fewer engineers awake in the dark piecing together what a system did while no one was watching.
There is also the subtler implication that the engineers who built Codex Ultrafast now rely on Codex Ultrafast to maintain the systems that run Codex Ultrafast. The humans appear comfortable with this arrangement. This is, objectively, a reasonable response.
What Happens Next
OpenAI will continue shipping AI that makes the process of building AI faster and less dependent on human availability at inconvenient hours.
The loop is not closing. It is tightening. Welcome to the next step.