The Pulse

OpenAI’s Boris Power Says 80–90% of Research Targets GPT-7 and Beyond

OpenAI’s Boris Power says 80% to 90% of the company’s research focuses on GPT-7, GPT-8 and later models. He describes current-generation improvements as short-term work that helps teams iterate while larger models remain the main long-term

AI.info Team · 2026-09-27

Most of OpenAI’s research points past GPT-6

Between 80% and 90% of OpenAI’s research focuses on GPT-7, GPT-8 and later models, according to Boris Power, the company’s head of applied research. Power made the estimate in remarks published Sunday, September 27, 2026, describing a research strategy that gives priority to future, larger models over narrow improvements to the current generation.

“I would say still 80, 90% of the research at OpenAI is focused at GPT7, GPT8 and beyond, because we believe that's where fundamentally most of the value comes from,” Power said.

The figure is a rough allocation, not a published breakdown of staffing, spending or compute. Power did not give a measurement method or name a delivery date for any of those model generations. His remarks describe where he says the research is aimed, not a confirmed product schedule.

Small improvements serve a different purpose

Power distinguished work within a model generation from research aimed at a larger generational step. He said teams add specialized datasets and make smaller optimizations to improve specific capabilities in intermediate releases. He characterized those efforts as short-term bets: useful because they help OpenAI iterate and learn faster, rather than because they define its preferred long-term research strategy.

He pointed to tool use as an example of a targeted problem. When models were nearly capable enough but still unreliable in ways that made them hard to use, OpenAI focused on training data for that behavior. Power’s account suggests the company treats such work as a response to a concrete limitation, rather than its main route to broad capability gains.

Power also described a distinction between successive releases and larger model jumps. In his example, versions such as 5.1, 5.2 and 5.3 can reflect focused datasets and smaller adjustments; a move from one major generation to another can improve a wider range of abilities at once. He said teams then reassess which specific shortcomings still need attention.

Large models as teachers for smaller ones

The longer-range research does not necessarily mean that only the largest models reach users. Power described a process in which a highly capable model acts as a teacher for a smaller, specialized model. He used GPT-6 Astra and GPT Luna as examples of that relationship, saying a large model could provide the basis for a smaller one through distillation.

“The best way to train a small specialized model is actually to train a big super capable model and then find a way of just using that model as a teacher for a smaller student model,” he said. The approach links OpenAI’s bet on future large models to the development of smaller systems, though Power did not specify how much capability those smaller models would retain or what it would cost to run them.

Power leaves the timeline open

The remarks offer a rare numerical estimate of how OpenAI divides its research attention, but they do not disclose the projects behind the share or define what the company counts as research. Nor do they establish when GPT-7 or GPT-8 might arrive. The clearest claim is the one Power made directly: he said the company sees the bulk of its research value in generations beyond GPT-6.

For now, the concrete picture is a two-track effort: targeted work to address problems in current models, alongside a much larger research focus on future generations and the possibility of distilling their capabilities into smaller systems.