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 is already aimed at GPT-7, GPT-8 and the generations after them, because that is where "most of the value" comes from. The Decoder reported his comments on September 27.

Small updates as intentional short-term bets

Improvements within a single generation, such as the step from GPT-5.1 to GPT-5.2, come from specialized training data and are intentional short-term bets. Power says they are seen inside the company as "extremely shortsighted". They help OpenAI iterate and learn faster today, but they are not the right long-term strategy. The real leap comes with a new model generation, when "everything else just works a lot better"; after each such jump, the company has to relearn where to invest to find quick wins.

The main challenge is onboarding, not model quality

According to Power, the biggest problem with today's AI assistants is not model quality but onboarding. Most ChatGPT users, he says, do not know what they can do with AI. Future models should get better at showing users what is possible and at anticipating what they need, instead of waiting for precise instructions.

From careful prompting to handing over a goal

The difference between generations is visible: GPT-4 required carefully written prompts, GPT-5 is easier to use but still needs a lot of feedback, and GPT-6 already works more like a capable colleague you can hand a goal to. For OpenAI, that means most of its research effort now goes into the generations that are still ahead.