official high credibility
Dark Factories

OpenAI Blog

A watershed moment in AI development. Codex 5.3 analyzed training logs and flagged failing tests during its own development.

Key metrics reported by OpenAI:

  • 25% faster development compared to prior iteration
  • 93% fewer wasted tokens during the build process
  • First frontier model described as “instrumental in creating itself”

This is the factory-level version of the dark factory concept applied to AI training itself — the model becoming an active participant in its own development.