The announcement emphasizes sustained execution and the ability for users to steer work while the agent is running.
What happened
OpenAI reports improvements on software-engineering and terminal benchmarks, alongside computer-use and knowledge-work results. It also says early model versions helped its staff debug training and inspect evaluation results.
Why it matters
An interactive agent can adapt when a user changes a requirement during execution. That capability matters for multi-step tasks where the initial prompt rarely captures every detail.
The bigger picture
Our take. Assistance with a model's development should not be confused with a system independently designing itself. For everyday users, the practical comparison is reliable task completion, including validation and the human effort needed to review changes.
Source checked November 18, 2026. Company claims are attributed; paragraphs marked “Our take” are editorial interpretation.