The remarks were made at the Fellows Forum AI Summit in late September 2026.
Power described improvements inside a model generation, such as moving from GPT-5.1 to GPT-5.2, as short-term work supported by specialized training data. He said OpenAI treats new generations as the larger step, after which the company must again identify areas for faster incremental gains.
He also said that onboarding, rather than raw model quality, is a principal constraint for current AI assistants. In his account, GPT-4 required careful prompting, GPT-5 requires less effort but still benefits from feedback, and GPT-6 works more like a colleague given a goal.
These are reported remarks, not an OpenAI product announcement. The source does not establish a GPT-7 or GPT-8 release schedule, technical specifications, availability, pricing, or that the stated research split will remain unchanged.