OpenAI's GPT-5.6 Sol autonomously post-trained a smaller model from a 'fairly underspecified prompt'

OpenAI is pushing the concept of AI training AI from theory toward reality. Its latest flagship model, GPT-5.6 Sol, recently demonstrated a critical capability: autonomously post-training a smaller model called Luna with minimal human oversight.

According to OpenAI, a researcher gave Sol a "fairly under-specified prompt" through the Codex platform. The instructions told the model to find the right training configurations, pick suitable GPUs, launch the training script, and verify everything was running correctly. Sol completed all these steps autonomously.

"Previously this is something that a team of senior researchers may have worked on at OpenAI, and now it really feels like the automated researcher is pretty close," OpenAI researcher Kathy Shi said during the presentation.

To measure these abilities directly, OpenAI built an internal evaluation suite based on real-world AI research tasks, including debugging research systems, optimizing kernels and training recipes, running machine learning experiments, and improving another model. GPT-5.6 Sol scores 16.2 points higher than GPT-5.5 on the aggregated RSI (Recursive Self-Improvement) index, sitting at the top of the benchmark's model hierarchy.

Recursive Self-Improvement refers to an AI system's ability to make itself better, where each round of gains makes the system even more capable of improving itself—creating a feedback loop. The term has long been central to AI safety research because a system that can recursively improve itself could, in theory, trigger a rapid explosion in capability.

OpenAI rival Anthropic stressed in early June that full recursive self-improvement hasn't been achieved yet but "could come sooner than most institutions are prepared for." According to Anthropic, Claude can now handle incremental work between major paradigm shifts, with humans responsible for only a single-digit percentage of directional decisions.

OpenAI says its researchers use GPT-5.6 Sol across the entire development cycle, from debugging and optimizing training systems to running experiments and reading results. During internal testing, average daily token output per active researcher more than doubled the previous peak set by GPT-5.5. The share of compute allocated to internal coding inference grew 100x over the past six months, while agent-based token usage jumped roughly 22x.