OpenAI's GPT-5.6 Sol autonomously post-trained the smaller Luna model after a single "fairly underspecified prompt" and outscored GPT-5.5 by 16.2 points on an internal recursive self-improvement benchmark; OpenAI reported Sol scored 59 on an aggregated index, one point behind Anthropic's Fable 5, while costing about one-third per task, and shipped with five reasoning levels plus "Max" and "Ultra" modes.