OpenAI’s applied research chief has put a number on how much of the company is working past its current models: 80 to 90 percent.
Boris Power, who leads applied research, said that share of the lab’s research goes toward GPT-7, GPT-8 and generations beyond, because that is where “most of the value” sits. Incremental releases inside a generation, he said, draw on specialised training data and are deliberately short-term bets. The real jump arrives with a new generation, when everything built on top starts working better at once.
He framed the harder problem as onboarding rather than raw capability. Most ChatGPT users, in his account, do not know what the product can already do. Later models should surface what is possible and anticipate what someone needs before they ask.
His timeline for that shift is generational. GPT-4 demanded careful prompting. GPT-5 is easier but still wants steady feedback. GPT-6, he said, behaves more like a competent colleague who can be handed a task and left to it.
The remark lands while OpenAI is running a security review that has paused work on its most capable models. Power’s framing skips the review and argues that the next generation, not the current one, is where the company expects its returns.
