Most of the research OpenAI does is aimed at models nobody can use yet. Boris Power, the company's Head of Applied Research, put the share aimed at GPT-7, GPT-8 and beyond at roughly 80 to 90 percent during a talk in Menlo Park on September 23. GPT is the family of AI models behind ChatGPT, so what he described is work aimed at generations that have not reached the public. The tension in the remark is the distance it draws between where a senior figure at OpenAI says the value is, and what the people using its product can get today.
A remark on an invite-only stage
Power was speaking at the Fellows Forum, a conference held on September 23 and 24, 2026 in Menlo Park and open by invitation only. The agenda lists him for a fireside chat titled Where Models Meet Reality, and his title there is Head of Applied Research. On stage, as reported afterwards, he said: "still 80, 90 of the research at OpenAI is focused at GPT-7, GPT-8, and beyond, because we believe that's where fundamentally most of the value comes from." The Decoder published its account of the talk on September 27 and attributes the quote to the conference's own recording.
The names carry part of the point. GPT-7 and GPT-8 describe the next generations of the family behind ChatGPT, the ones beyond what the product runs on today.
The figures deserve care. "80, 90" is how the numbers left his mouth: two round values offered aloud, not a measured share with a method behind it. The Decoder attributes the remark to Power in its report, and no OpenAI post, release or transcript carries the 80 to 90 percent claim. Read as a statistic, the number would say OpenAI has all but stopped improving the tools people use. Read as what it is, an estimate of where effort is concentrated, it says something narrower: that the bulk of the research is pitched at a model generation beyond the current one.
Why the effort points past what exists
His reason was short: that is where he believes most of the value comes from. The nearer horizon got less deference. Asked about the incremental steps within a generation, the kind of change that moves a model from version 5.1 to 5.2, Power said those investments are seen inside the company as "extremely short-sighted". OpenAI makes them because they help it iterate and learn faster today, he said, not because they are the right long-term strategy. As The Decoder describes them, these within-generation improvements are intentional short-term bets.
That version number needs unpacking. A name like GPT 5.2 marks a revision to a model that already exists rather than a new generation. A new generation arrives under a new name with a fresh set of capabilities; a point release is the smaller, more frequent update to something already in use. The two kinds of work reach a user differently. One is a step change in what the product can do, the other a run of smaller improvements that arrive between the big announcements.
If Power's account holds, the split has a practical shape. The improvements a user notices between generations come from the line of work he calls a short-term bet, while the bulk of the research effort goes somewhere they cannot yet reach. That is an inference from his description of the split, not something he said in those words, and he did not claim OpenAI is neglecting the models it ships.
The claim about what actually holds users back
The Decoder reports a second view from the same talk, separate from the research share. In its account, Power does not see model quality as the biggest problem. Most people using ChatGPT do not know what they can do with it, and future models should get better at showing what is possible. The same account describes a progression in how the tools take a request: GPT-4 needed careful prompting, GPT-5 is easier to use but still calls for a lot of feedback, and GPT-6 works more like a capable colleague you can hand a goal to.
Those lines rest on The Decoder's reading of the recording it cites, not on a published transcript. The first part is Power's assessment of ChatGPT's audience; the second is his description of how the tools have changed as they moved from one generation to the next. The onboarding point is also a judgment about the people using the product, offered by someone who helps build it.
Where that leaves anyone using ChatGPT
The two claims sit awkwardly together. If most users do not know what the tools can already do, a more capable model does not by itself close that gap. The problem Power names is onboarding, which is about how the product introduces itself and what it shows a new user, not only how good the model underneath is. On that account the limit is not just what the software can do but what people find out it can do.
And if 80 to 90 percent of the research effort is aimed past the current generation, then the near-term experience of ChatGPT is shaped mostly by the shorter-term work he placed lower in the company's priorities. The remarks attach no date to the bet. Nothing said on that stage fixes when GPT-7 will exist, what it will do, or how the company will know it has arrived.
The two halves of the talk describe two clocks running at once: the steady flow of updates to what already exists, and a slower effort aimed at models OpenAI has named but not released. For anyone using ChatGPT today, the practical consequence is narrower than the headline number sounds. The improvements they see soon are likely to come through the releases Power called a short-term bet, while the value he points to waits on models that do not exist yet.