OpenAI Details Reasoning Levels in GPT-5.6 Sol
Vaibhav Shrivastav, OpenAI’s Head of Ecosystem & Infrastructure for EMEA, explained the new reasoning gradation in GPT-5.6 Sol, which differs from GPT-5.5—users should start one level lower than they’re accustomed to. Light and Low are for fast, simple tasks; Medium suits planning and analysis; High and xHigh target complex multi-step problems and verification. Max gives the model more time on a single task, while Ultra parallelizes work across multiple sub-agents. Shrivastav advises starting low and scaling up only when needed, as higher levels consume more time and tokens—a nuanced reality that stands in contrast to Greg Brockman’s stated goal at the Big Technology AI Summit last month: making ChatGPT as simple as possible.