OpenAI has introduced a gradation of reasoning levels for its new GPT-5.6 Sol model, allowing for flexible adjustment of analytical depth based on specific tasks.
What Happened
A system of complexity levels has been introduced for the GPT-5.6 Sol model: Light and Low for quick tasks, Medium for planning tasks, and High and xhigh for complex multi-step processes. Specialized modes have also been presented: Max, which increases the time spent on a single task, and Ultra, which utilizes the parallel work of several sub-agents. When using the new version, it is recommended to start with one level lower than usual.
Context
This initiative marks a transition from fixed computational complexity to a granular resource management model (compute-on-demand). Instead of paying for abstract "general intelligence," developers receive a tool to balance cost, latency, and response quality.
Why It Matters for the Industry
Introducing controlled computational resource consumption through reasoning levels allows for the optimization of API cost and response speed. This makes complex agentic systems more manageable and changes the economics of building AI agents, allowing for precise targeting of complexity to specific tasks and optimization of unit economics.
Why It Matters for Users
Users will be able to use GPT-5.6 more efficiently by selecting the minimum necessary complexity level. This will allow for saving tokens and time without sacrificing quality where high reasoning depth is not required.
What Is Not Yet Known / Limitations
Legal specialists (IP/Privacy Counsel, EU Regulatory Counsel) point to new risks associated with data management and the transparency of decision-making chains in multi-agent systems.
Sources
Author
Look at AI, Editorial Team