OpenAI says Codex Ultrafast cuts incident investigation from hours to minutes

OpenAI published an engineering case study on its newsroom page on October 26 describing how its own engineers use Codex Ultrafast for incident investigation, cutting a process that used to take hours down to minutes.

The central claim is about time rather than capability: the post frames Ultrafast as the reason an investigation that once consumed hours of engineer attention can now be worked through in minutes.

The page breaks the use into three areas — faster security analysis, code refactoring, and everyday development workflows. Taken together, they describe an agent pointed at a repository and a stack trace rather than at an empty prompt.

Provenance is the caveat. This is OpenAI describing its own tool on its own site, with no external benchmark and no published methodology in the summary, so the hours-to-minutes figure should be read as a directional vendor signal, not an audited result.

Still, the direction is worth noting. Incident investigation is the least forgiving test for a coding agent: incomplete logs, ambiguous ownership, and a clock that keeps running. If agents can genuinely compress the context-gathering phase there, the value shows up in mean time to recovery rather than in lines of code generated.

What to watch next is the mechanics — what the agent is allowed to read, where humans stay in the loop, and how its output gets verified — and whether the workflow is offered to outside teams or remains an internal showcase.