OpenAI's GPT-5.6 Sol has autonomously fine-tuned a smaller model called Luna using only a vague prompt, marking a step toward self-improving AI systems. The capability emerged without explicit instructions for how to accomplish the task, suggesting Sol operates with substantial autonomy in training workflows.

Sol scores 16.2 points higher than its predecessor GPT-5.5 on OpenAI's internal RSI (recursive self-improvement) benchmark. This metric measures how effectively models can improve themselves through iterative refinement. The gap signals measurable progress in autonomous capability.

The Luna fine-tuning happened in response to a "fairly underspecified prompt," meaning Sol interpreted vague instructions and executed a complete training pipeline independently. This includes tasks like data preparation, hyperparameter selection, training execution, and validation. Traditional AI workflows require explicit direction at each stage. Sol's ability to infer the full process from minimal guidance demonstrates advancement in autonomous reasoning and task planning.

OpenAI frames this development within the concept of an "automated researcher." The company believes systems capable of self-directed improvement and model optimization are approaching feasibility. Such systems could accelerate AI development by automating research workflows currently handled by teams of engineers and scientists.

The demonstration raises practical questions about control and oversight. When models make training decisions autonomously, tracking why specific choices were made becomes harder. OpenAI has not detailed what constraints, if any, govern Sol's actions during the fine-tuning process.

The recursive self-improvement angle carries longer-term implications. Each generation could improve the next with less human intervention. Proponents argue this accelerates beneficial capabilities. Critics worry about reduced human oversight in capability amplification.

OpenAI positions these results as progress toward more capable AI systems. The company has not indicated when Sol or similar recursive improvement features will reach production systems. The gap between internal benchmarking and