GPT-7 hasn't shipped yet, but most of OpenAI's research firepower is already pointed at it. On September 27, 2026, Boris Power, OpenAI's Head of Applied Research, said in an interview that roughly 80 to 90 percent of the company's research effort is going toward GPT-7, GPT-8, and whatever comes after — because "most of the value" comes from generational leaps, not point releases within a generation.
According to tech outlet The Decoder, Power described updates like GPT-5.1 to 5.2 as deliberate short-term bets: they deliver improvements through specialized training data and help the team learn fast, but inside the company such incremental updates are seen as "extremely shortsighted." In his view, the real leap only happens when a new generation lands — "everything else just works a lot better" — and after each jump, the team has to relearn where the quick wins are.
Why keep shipping "shortsighted" updates
Power isn't arguing to kill point releases; he's describing a resource-allocation priority: generational leaps set the ceiling, point releases only set how fast you climb toward it this generation. For OpenAI, the value of 5.x iterations is "learning faster today," not "building the future." It also explains why users keep feeling that small updates are incremental by design — inside the company, they were never the main event.
Each generation gets easier to live with
Power also made a point that has nothing to do with model architecture: the biggest bottleneck for today's AI assistants isn't model quality, it's onboarding. Most ChatGPT users don't know what AI can actually do for them. GPT-4 demanded careful prompting; GPT-5 is easier to use but still needs a lot of feedback; GPT-6 already works like "a capable colleague you can just hand a goal to." Future models should surface what's possible and anticipate user needs — exactly the kind of problem only a generational leap can solve systematically.
What developers and users should make of it
For developers, the practical takeaway is direct: don't plan product roadmaps around 5.x point updates as if they were the capability ceiling. They're maintenance, not signal; the real variable is the next generational jump. OpenAI just pushed GPT-6 Sol and Luna toward everyday tasks (OpenAI launches GPT-6 Sol and Luna: API prices halved, aimed at everyday tasks), yet its research focus is already past them — a sign that the company's compute and talent bets are on a more distant future. Rivals aren't standing still either: Google's Gemini 4 has entered early post-training (Gemini 4 enters early post-training, Google's new chief says release well before year-end). For ordinary users, the conclusion is simpler: stop chasing every point release, and wait for the next generation.
