Choosing GPT-5.6 Sol, Terra, and Luna in Codex

Codex's three models

Sol is a model tailored for ambiguous and difficult tasks, and it excels at complex problems where deep investigation and polish can significantly change the outcome. It connects ideas, catches easily missed details, and can offer unexpected insights.

Terra is an all-rounder suited for everyday implementation, testing, and multi-step tasks. It handles ambiguity, finds context, and coordinates sub-agents effectively, while tending to converge on solid results rather than chasing every detail.

Luna is a fast option for clear, well-scoped tasks. It's well suited for high-volume workflows like extraction, classification, conversion, and structured summarization, and at the xHigh reasoning level it can produce high-quality results even on scoped implementation tasks.

Ultra and prompt writing

Sol Ultra provides top-tier intelligence for the most difficult tasks. It maximizes reasoning and combines it with multi-agent collaboration, allowing agents to investigate deeply from multiple directions at once. Token usage is considerable, so it's best used only when depth and coordination are worth the cost.

An effective prompt should provide direction, not a journey. It's important to include four elements: the goal, context (code, documentation, Slack threads, issues), the deliverable and its boundaries, and completion criteria.

Model selection strategy

It's recommended to use Sol Medium as a baseline. Generally, smaller models require higher reasoning levels, so a task suited to Sol Medium might require Terra High or Luna xHigh. For high-stakes tasks or scattered context, Sol Ultra is effective; for scoped tasks with some complexity, Terra High works well; and when speed matters, Luna xHigh is the effective choice.