Codex for moonshots and everything in between
Some missions demand deep planning and coordination. Others are a straight shot. The same is true of the work you hand to Codex, which is why GPT-5.6 gives you three models to choose from. If you’re unsure where to start, go with Sol Medium.
Where each model shines
Sol: for complex, open-ended work — Sol is built for ambiguous, difficult, or high-value work where deeper investigation and polish can change the outcome. It connects ideas across a problem, catches details that are easy to miss, and can surface useful insights you may not have thought to ask for. That depth can also make Sol more efficient on hard debugging tasks, where avoiding a few wrong turns is worth far more than a faster first attempt.
Terra: the pragmatic all-rounder — Terra is well suited to everyday implementation, testing, and multi-step work that still requires good judgment. It handles ambiguity, finds relevant context, and coordinates subagents effectively, while tending to converge on a solid result without pushing for every last detail or insight. Terra High is particularly useful when the scope is understood but the implementation still has meaningful complexity.
Luna: for clear, well-scoped work — Luna is a fast option, which makes it a natural fit for high-volume workflows like extraction, classification, transformation, and structured summaries. It can also take on more substantial implementation work when the scope and expected outcome are clear. At xHigh reasoning, Luna can deliver high-quality results on bounded implementation work.
What Ultra means
Most tasks won’t need Ultra. For the hardest work, Sol Ultra brings the highest level of intelligence available in Codex, combining maximum reasoning with proactive multi-agent collaboration. Agents can investigate deeply while making progress across multiple lanes at once. It uses significantly more tokens, so save it for work where that added depth and coordination are worth it.
Planning is a great use case. With the right plugins, you can point Codex at a Slack thread, relevant GitHub issues and PRs, docs, code, and git history. Ask Ultra to pull that context together, work through the ambiguity, and produce a clear implementation plan.
Once the scope is defined, implementation becomes much easier to hand off to Sol Medium or High, Terra High, or Luna xHigh.
Large plans do not always need Ultra. Sol Medium can produce strong results too, especially when you ask it to use subagents proactively and split the work across clear lanes. Save Ultra for when the stakes, ambiguity, or amount of context justify the extra depth.
Out of the box, Codex creates subagents that inherit the conversation so far and use the same model family and reasoning level as the parent agent. Those defaults are deliberate, and the model knows how to use them effectively. Coming later this week, you’ll be able to customize those choices through skills or prompts, using lighter settings for context gathering while keeping stronger defaults for implementation.
Give Codex a clear finish line
The best prompts give Codex a direction, not an itinerary. Sol can discover context from the tools available to it, follow promising leads, and work through ambiguity without being told every step. What it needs from you is a clear outcome, a few good places to start, and a way to recognize when the work is done. If the problem spans multiple lanes, ask it to bring in subagents early. A useful prompt covers four things:
- Goal: the outcome you want and who it needs to work for.
- Context: the code, docs, Slack threads, issues, or other starting points that can help Codex understand the problem.
- Output and boundaries: what Codex should produce, what should stay untouched, and where it needs approval.
- Finish line: the checks, evidence, or decisions that make the result ready to hand off.
Example Sol Ultra planning prompt
“Start from
this
Slack thread. Find the related issues, PRs, docs, code, and git history, then turn what you learn into a clear implementation plan. Call out the scope, approach, risks, open decisions, and how we’ll know it works. Present the plan as a clean, self-contained HTML page we can review together. Don’t implement yet.”
Match the effort to the work
Sol Medium is a useful baseline for tuning effort. A good rule of thumb is to increase reasoning as models get smaller, so a task suited to Sol Medium may call for Terra High or Luna xHigh.
If you want to match the model more closely to the task, here are a few good alternatives:
- Sol Ultra for high-stakes work, scattered context, or problems that are still taking shape.
- Terra High for well-scoped implementation that still has meaningful complexity.
- Luna xHigh for well-scoped implementation where speed matters.
Once you know what needs to be built and why, everything else gets easier. You can hand off clear pieces of work, pick the right model for each, and spend less time course-correcting along the way.
