OpenAI Explains GPT-5.6 Sol's Seven Reasoning Levels, From Light to Ultra
Written by
NextStair
OpenAI employee Vaibhav Srivastav has detailed how GPT-5.6 Sol's reasoning tiers work, from quick single-step tasks at Light all the way to Ultra, where multiple subagents run in parallel on different parts of a complex problem.
A Reasoning Slider, Not a Simple On-Off Switch
GPT-5.6 Sol does not operate at a single fixed intelligence level. It ships with seven distinct reasoning settings that users can dial up or down depending on what a task actually requires.
OpenAI employee Vaibhav Srivastav laid out the full breakdown on X on July 10, 2026, offering the clearest official guidance yet on when to use each tier and what separates them under the hood.
The Seven Levels Explained
Light and Low sit at the bottom of the stack and are built for quick, well-scoped tasks where the answer is clear and speed matters more than depth.
Medium steps up to handle planning, analysis, or tasks with a few moving parts that require some reasoning but not sustained effort.
High and xhigh are for harder multi-step work involving tradeoffs, research, or careful verification, the kind of tasks where missing a step causes compounding errors downstream.
Max and Ultra operate differently from the tiers below them. Max keeps a single model working on one hard problem for longer, giving it more time to reason through a difficult question before committing to an answer. Ultra goes further by deploying multiple subagents in parallel, each handling a different part of a larger task simultaneously.
Cost and Speed Scale With the Setting
Higher reasoning levels take more time and consume more tokens. Srivastav's core recommendation is to start at the lowest setting that gets the job done and scale up only when the result actually needs more thinking.
The approach matters for cost control. Running every query at Ultra burns significantly more tokens than running lighter tasks at Light or Low, and for most everyday use cases the difference in output quality does not justify the extra spend.
Switching From GPT-5.5 Requires a Recalibration
Srivastav flagged one important note for users migrating from GPT-5.5. The reasoning levels do not map directly between the two models. A task that ran well at High on GPT-5.5 may not need the same setting on GPT-5.6 Sol.
His advice is to try starting one level lower than your usual setting on a familiar task, evaluate the result, and increase only if needed. The change reflects differences in how the two models process reasoning effort rather than any capability regression.
Why This Design Matters
The tiered system reflects a broader shift in how frontier AI labs are thinking about model deployment. Rather than offering one model that tries to do everything at the same cost, OpenAI is letting users trade compute against complexity at the task level.
For businesses running high-volume automation workflows, the ability to route simple tasks to Light and reserve Ultra for genuinely complex multi-step work translates directly into lower API costs without sacrificing quality where it counts.
The same design decision contributed to the Mac file deletion incident reported on July 11, however. Matt Shumer was running GPT-5.6-Sol on Ultra mode at the time, and critics have since questioned whether deploying multiple autonomous subagents in parallel on a machine with full filesystem access raises the risk of destructive commands executing faster than any human oversight can intervene.