OpenAI directs 80–90% of research toward GPT-7 and beyond

Two kinds of research bets

Power says each new generation can bring more noticeable improvements and make everything around it work significantly better. But after a leap, OpenAI has to reassess which investments will pay off quickly.

Changes between versions of the same model, such as GPT-5.1 to GPT-5.2, typically rely on specialized training data. OpenAI treats these as short-term bets. They help the company iterate faster and learn from current results, even as the approach is seen internally as excessively short-sighted.

The bottleneck may be the user

For Power, the main constraint on AI assistants is not model quality but whether people know what to ask them to do. He says most ChatGPT users do not know what they can delegate to AI. Future models, in his view, need to make their abilities clearer and anticipate users’ needs.

The progression he describes is from prompting to delegation: GPT-4 needed carefully crafted prompts; GPT-5 is easier to use but still requires substantial feedback; GPT-6, he says, is more like a capable colleague that can be given a goal.

I think that makes the 80–90% figure more revealing than a simple forecast about model releases. OpenAI is putting most of its research behind the idea that a new generation can change what users get from the whole system—not just produce a better answer. What the allocation does not tell us is how much of that work is aimed at helping users discover what to delegate, rather than making the models themselves more capable. If that distinction holds, the next leap will be judged as much by whether people know how to use it as by what the model can do.