OpenAI employee Vaibhav Srivastav has published a guide detailing how the five reasoning tiers of GPT-5.6 Sol correspond to specific task complexities. The breakdown assigns “Light” and “Low” settings to simple, clear-cut jobs, while “Medium” suits planning and analysis. More demanding work requiring careful verification or multi-step logic falls under “High” and “xhigh”. The top two tiers, “Max” and “Ultra”, operate differently by either extending time on a single problem or deploying parallel sub-agents. Srivastav advises users to begin with lower settings and only increase resources when necessary, noting that these levels do not align with GPT-5.5’s previous structure. Anyone switching models should start one tier lower than their current habit.
This documentation does not resolve the broader issue that Sol’s Pro tiers remain unavailable despite earlier leaks in a genomics benchmark paper. The system still requires manual benchmarking to select the correct configuration, contradicting OpenAI’s aim for a nearly interface-free experience. The setup may primarily serve to gather usage data rather than simplify the user journey. Until the Pro tiers launch, ambitious users will continue to struggle with optimisation.
- Higher tiers consume significantly more tokens and time.
- Pro tiers are still missing from the current release.
- Users must run their own benchmarks to find the right level.
