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

I'll pull the linked thread and HN item for verified facts, then draft the paragraphs from only what's there.Eric Provencher, who works on Codex DX at OpenAI, published a practical guide to picking among GPT-5.6 Sol, Terra, and Luna inside Codex. The post frames Codex as a place for both moonshot work and routine delivery, and it treats the three models as tools matched to mission shape rather than a single default for every task. If you are unsure where to start, the guidance is direct: use Sol Medium. On Hacker News the linked item had 1 point and 0 comments at capture, so the primary source is the article itself rather than community debate.

The product mechanics are a tiered model family plus reasoning levels, not a one-size agent. Sol is positioned for complex, open-ended, or high-value work where deeper investigation and polish change the outcome, including hard debugging where avoiding wrong turns matters more than a fast first attempt. Terra is the pragmatic all-rounder for everyday implementation, testing, and multi-step work that still needs judgment; Terra High is called out when scope is clear but implementation remains meaningfully complex. Luna is the fast path for high-volume extraction, classification, transformation, and structured summaries, and at xHigh reasoning it can still handle bounded implementation. Sol Ultra is reserved for the hardest cases: maximum reasoning with proactive multi-agent collaboration across multiple lanes at once, at significantly higher token cost. Out of the box, Codex subagents inherit the conversation and use the same model family and reasoning level as the parent; customization through skills or prompts was described as coming later that week.

For engineers and builders, the useful shift is routing work by ambiguity and stakes, not by habit. Planning is the clear Sol Ultra (or strong Sol Medium with proactive subagents) use case: with the right plugins, point Codex at a Slack thread, related GitHub issues and PRs, docs, code, and git history, and ask it to produce a self-contained plan before anyone implements. Once scope is defined, handoffs go to Sol Medium or High, Terra High, or Luna xHigh depending on how bounded and speed-sensitive the remaining work is. Provencher’s prompt pattern is concrete: state the Goal, starting Context, Output and boundaries, and a Finish line with checks or evidence of done. Give direction and a clear outcome, not a full itinerary, and ask for subagents early when the problem spans multiple lanes.

Competitive and market context is implicit in the design: Codex is productizing model choice as part of the coding agent workflow, with named tiers (Sol, Terra, Luna) and effort knobs (Medium, High, xHigh, Ultra) instead of a single opaque “smartest model” button. That sits in the same product space as other agentic coding stacks that already force tradeoffs among depth, cost, latency, and parallelism. The article does not publish benchmark tables or API price sheets; it sells selection rules and workflow shape. Relative to a flat “always use the flagship” policy, the pitch is operational: match effort to work so teams spend less time course-correcting and less budget on Ultra-grade runs that did not need them.

Practical takeaway: treat Sol Medium as the baseline, then step effort up as the model tier steps down—so work that fits Sol Medium may map to Terra High or Luna xHigh when you want the smaller model. Reach for Sol Ultra when stakes, ambiguity, or scattered context justify the token burn; use Terra High for well-scoped but still complex implementation; use Luna xHigh when the job is well-scoped and speed matters. Watch next for subagent customization (lighter settings for context gathering, stronger defaults for implementation) and for whether your own workflows stabilize into a small set of pairings—planning on Sol Ultra or Sol Medium, execution on Terra High or Luna xHigh—rather than a single global default.