OpenAI Says Its Models Are Now Running Research That Improves the Next Models

OpenAI's September 8 essay, The Work Now Within Reach, makes an economic argument rather than a capability one. Better models plus falling cost per task expand the set of work worth automating, and revenue from that expansion funds the next training run and the next data center. The company cites deployed examples, including customer service handling in which about 65 percent of contacts are resolved without a human.

OpenAI has also described its own models doing research work. According to the company, GPT-5.6 Sol improved a draft model by designing and running hundreds of experiments on its architecture, varying its size, structure and features. Separate reporting from The Decoder describes the same system post-training a smaller model from a loosely specified prompt. AI Post summarized the claim for its audience as AI already helping build its successor, noting that humans supervised tasks that would otherwise take a skilled researcher several days.

This description deserves scrutiny. Automating hyperparameter sweeps and ablation studies is real engineering leverage, and it is also the part of research that was already closest to automation. Neither OpenAI nor an outside evaluator has published a controlled comparison showing how much of a capability gain came from model-run experiments rather than from more compute and better data.

Greg Brockman, OpenAI's president, said at the GPT-6 Astra launch that artificial general intelligence is arriving in pieces rather than in one moment, a notably softer claim than the timelines circulating around it.