OpenAI says Codex cut custom code by about 90% in FDE projects

What is changing?

OpenAI explained on October 22 how its Forward Deployed Engineering team closes the “deployment gap” between model capability and reliable production use. The team defines business outcomes and evaluations with customers, then gives Research concrete capability gaps when current models fall short.

At one semiconductor customer, more than 10 verification workflows reached production within three to six months. OpenAI reports about 48% time savings from AI-based directed test generation, a 22% median performance improvement from kernel optimization, and debugging cycles compressed from weeks to hours.

The custom applications were later rebuilt around Codex. Skills, plugins and MCP connections replaced separate interfaces and tools, cutting custom code by roughly 90% per project. A deployment that took two months for the first team was extended to a second team in one day.

OpenAI frames FDE as a way to turn individual engagements into repeatable product and model capabilities. These results are provider-reported examples; other organizations will need their own evaluations and production measurements.