OpenAI employee Vaibhav Srivastav detailed how GPT-5.6 Sol's five reasoning levels correspond to varying task complexities. According to Srivastav, the 'Light' and 'Low' levels are designed for quick, straightforward tasks, while 'Medium' handles planning and analysis. 'High' and 'xhigh' are for complex, multi-step work requiring careful verification. 'Max' and 'Ultra' operate differently, with 'Max' allowing models to spend more time on a single problem and 'Ultra' deploying multiple sub-agents in parallel to tackle different parts of a task. Higher levels consume more time and tokens, and Srivastav advises starting at a lower level and scaling up only when necessary.
Srivastav noted that these levels do not align with GPT-5.5's tiers, and users transitioning from previous versions should start one level lower than they were accustomed to. He also mentioned that the setup might help OpenAI gather usage data, though ambitious users may still struggle to select the right level without running their own benchmarks. The Pro tiers for Sol remain unavailable, as they were previously leaked in a genomics benchmark paper.
Source: thedecoder
