OpenAI Says 80 to 90 Percent of Its Research Targets GPT 7 and Beyond

Boris Power, OpenAI's Head of Applied Research, says 80 to 90 percent of the company's research effort already goes toward GPT 7, GPT 8, and beyond. He disclosed the figure in comments on how OpenAI allocates its research resources, offering a rare look at how one of the industry's largest AI labs splits its work between shipping near-term model improvements and placing long-term bets on future generations.

The remaining 10 to 20 percent of research, by implication, covers work on the current generation of models. Power said improvements within a single generation are intentionally short-term bets. The framing suggests OpenAI treats each released model line as a fixed foundation, with only marginal refinements until the next major version arrives, while the bulk of the organization's capacity builds the architectures, training methods, and capabilities that will define GPT 7 and GPT 8.

That allocation carries weight beyond OpenAI's own roadmap. Frontier AI development is increasingly concentrated in a handful of labs, and where those labs point their research budgets shapes what capabilities arrive in the market — and in the hands of users — over the next several years. A figure of 80 to 90 percent aimed at models two or more generations out signals that OpenAI expects the biggest leaps in AI capability to come from future model families rather than from iterating on what it has already shipped.

Power also made a second, arguably more striking claim: he does not see model performance as the biggest problem facing AI today. Instead, he pointed to a gap on the user side. Most users don't even know what they can do with AI, he said.

That assessment reframes a common industry debate. Much of the public conversation around AI centers on benchmark scores, reasoning ability, and the race toward more capable systems. Power's comment suggests that, at least from OpenAI's applied research vantage point, the constraint on AI's impact is not raw model capability but user awareness and adoption — the gap between what current models can do and what the average person understands they can do.

The two statements fit together. If most users underutilize existing models, the marginal value of squeezing incremental performance out of the current generation is limited, which supports shifting the research portfolio toward future generations. The next major capability jump, in Power's framing, matters more than polishing the present one.

The comments come from Power's role leading applied research at OpenAI, the division concerned with translating research advances into real-world usefulness. That position gives his claim about user awareness particular relevance: applied research teams sit closest to how people actually interact with the company's models.

For OpenAI, the disclosure also serves as a signal to competitors, customers, and developers. Committing the large majority of research capacity to GPT 7, GPT 8, and beyond tells the market that OpenAI expects sustained multi-generation investment rather than a plateau at current capability levels. For businesses building on OpenAI's models, it hints that significant capability changes will arrive on a generational cadence, with the current lineup receiving only deliberate, short-term refinements in between.

Power's observation about users carries implications for the broader adoption curve as well. If the bottleneck is knowledge rather than performance, then education, product design, and interface work may determine how quickly AI delivers economic value — a task that falls to more than just model developers.

OpenAI has not specified what GPT 7 or GPT 8 will deliver or when they will arrive. What Power has made clear is where the company's research weight sits: on the generations that do not exist yet, and on a bet that the future models — not better versions of today's — will define what AI can do.