GPT-5.6 Sol scored 49.8 Pass@1 points per million tokens on APEX-Agents, making it the most token-efficient frontier model, roughly 1.5 times more efficient than its competitors. In its maximum reasoning and pro mode configuration, Sol achieved a 40.0% Pass@1 score, closing within 3.3 points of Claude Fable 5, which leads at 43.3%. The model's agent style favors many light steps, using around 28,000 tokens per tool call at xhigh, compared to 63,000 for Claude Opus 4.8 and 71,000 for Fable 5.