Codex 6 Sol: Ultra Reasoning Arrives in the CLI

Ultra in the CLI is more than boosted reasoning

What stands out here is not just another toggle, but a change in how the system works: Codex 6 Sol now offers six reasoning levels directly in the CLI. OpenAI’s official GPT-6 Sol documentation lists Low, Medium, High, Extra high, Max, and Ultra. As of September 23, 2026, Medium is listed as the default.

OpenAI’s reasoning API documentation separately distinguishes the reasoning mode from reasoning effort. Choosing a model, therefore, does not automatically mean maximum task processing depth, while leaving the setting unspecified keeps the system at a medium level. That matters for configuration and reproducibility: the same request to the same model can run with a different reasoning budget.

The most interesting change begins with Ultra. It is described as maximum reasoning with automatic task delegation, so this is no longer only about a longer internal search for an answer. Conceptually, it is closer to an embedded multi-agent loop in which the main model can distribute work across subtasks.

The first signal of the feature came from user observations of Codex CLI, while the documentation explains what sits behind the new interface option. It does not, however, reveal orchestration details: how many subtasks are created, how context moves between them, or how the final result is assembled.

Complex work changes the workflow itself

The practical shift is straightforward: developers can now choose not only the model, but also the form of computation applied to a task. Medium looks like the baseline for everyday coding, while High, Max, and Ultra target more difficult debugging and multi-step repository work.

I would focus first not on the Ultra label, but on the cost of coordination. Automatic delegation can improve decomposition, yet it can also introduce duplicated work, greater context use, and mistakes when partial results are combined. Without transparent tracing, it is hard to tell whether the model divided the task intelligently or simply generated more computation.

This is not an empty cosmetic update: reasoning is gradually becoming part of the standard developer interface rather than a hidden API parameter. But Ultra’s real value will depend not on maximum reasoning depth, but on how reliably Codex coordinates its own subtasks. That is the line between a useful agentic mode and an expensive illusion of productivity.

We previously covered Codex’s rollout in ChatGPT for Android and what it means for remote development teams. That context helps frame how a new reasoning mode could change Codex’s practical role in engineering workflows.