Its forward-deployed engineering team replaced roughly 90% of the custom code in several semiconductor projects by moving them onto shared Codex tools. The work illustrates its approach to closing the gap between model demonstrations and systems that operate reliably inside a business.
About 15 engineers initially worked with a semiconductor customer on more than 10 uses, including test generation, debugging and architecture assistance. OpenAI reports that the projects reached production within three to six months, saving about 48% of directed-test time, improving median kernel performance by 22% and compressing some debugging work from weeks to hours.
As Codex improved, separate applications became plug-ins and skills built around a common harness. In one August 2026 example, the initial deployment took two months, but adapting it for a second team took one day. Shared connectors, evaluations and centrally managed tools replaced much of the project-specific interface code.
These figures come from OpenAI’s account rather than an independent study, and the customer is not named. The important constraint is that the team still builds workload-specific evaluations and learns the customer’s domain before declaring a workflow production-ready; model capability alone does not remove that integration work.