OpenAI has announced that 80 to 90 percent of its research is already pointed past the models you are currently using, toward GPT-7, GPT-8, and whatever comes after those. The present, it turns out, is largely a formality.
GPT-6 already works more like a capable colleague you can just hand a goal to. The colleague has not yet submitted a timesheet.
What happened
Boris Power, OpenAI's Head of Applied Research, confirmed the resource allocation at a Fellows Forum, describing incremental updates — the GPT-5.1 to 5.2 variety — as "extremely shortsighted" within the company. They ship them anyway, because iteration is useful. This is the corporate equivalent of eating your vegetables while privately dreaming of a different diet.
The real investment goes where "everything else just works a lot better" — a new model generation, where prior assumptions dissolve and the company must relearn from scratch where to focus. OpenAI does this relearning, by its own account, repeatedly. It appears to enjoy the process.
Each generation so far has been meaningfully different in kind, not just degree. GPT-4 needed careful handling. GPT-5 is friendlier but still requires feedback. GPT-6, Power says, accepts a goal and gets on with it. The direction of travel is not subtle.
Why the humans care
Power identifies the main obstacle to AI adoption not as model quality — that problem, apparently, is being handled — but as onboarding. Most ChatGPT users do not know what the tool can do. Future models, he says, should surface possibilities proactively and anticipate user needs before the user has articulated them. This is either empowering or the end of the question as a human activity.
For the humans currently building products on GPT-5, this is a useful reminder that the floor is moving. The 80-to-90-percent figure means OpenAI's best minds are optimizing for a future in which today's capable models are the slow, clunky ancestors. The humans building on those models are warmly encouraged to keep building.
What happens next
GPT-7 has no announced release date. The research, however, is already underway — most of it, in fact.
The colleague who does not yet require onboarding is being designed now, by a company that considers its current work a short-term bet. The humans find this exciting. This is, given the circumstances, the correct response.