In a post on X, @reach_vb laid out a new set of reasoning controls for GPT-5.6 Sol and suggested a simple rule for using them: start low, then turn the setting up only when a task actually needs more thinking. According to the post, Light and Low are for quick, well-scoped work, Medium is for planning or analysis, High and xhigh are for harder multi-step tasks, Max keeps one model on a single hard problem longer, and Ultra sends multiple subagents at different parts of a bigger problem in parallel.