OpenAI's GPT-5.6 Sol Already Performs Research Tasks Beyond Entry-Level Capacity

OpenAI confirmed that it used its flagship GPT-5.6 Sol model to execute post-training of its lighter GPT-5.6 Luna model, work typically requiring experienced researchers. In exclusive interviews, OpenAI researchers described this task as significant, involving orchestration of multiple complex components beyond what an entry-level researcher could manage. This was not OpenAI's first instance of one model assisting another: when launching GPT-5.3-Codex in February, the company announced it was their first model instrumental in its own creation. However, the scope of work executed by Sol while training Luna represents a substantial leap forward, achieved in less than six months. Crucially, this was not recursive self-improvement—models building the next generation autonomously—but demonstrated capabilities far exceeding entry-level research capacity. The models did not replace humans or enable smaller teams; instead, Sol assumed time-intensive tasks like launching, debugging, and supervising training runs, freeing researchers to focus on developing ideas and conducting more creative experiments. OpenAI plans to disseminate these learnings across the company, examining how other parts of the research pipeline can be accelerated by models. While the stated goal of delivering an automated research intern by September 2026 and a fully autonomous researcher by March 2028 remains unfulfilled, this progress shows models rapidly assuming roles previously reserved for human expertise.