official high credibility
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.