# Choosing GPT-5.6 Sol, Terra, or Luna in Codex — Eric Provencher ## 🗂️ Triage — annotated highlights surfaced > [!note]- Cross Reference > - **Increase reasoning level as model size decreases** — the core substitution heuristic; directly comparable to Claude's model+effort matrix. #xref #aphorisms > - **Most tasks don't need the top-tier Ultra mode** — vendor restraint against reflexive escalation; mirrors Courtney's credit-routing discipline. Connects to [[AI Cost & Tokenomics MOC]]. #xref #addendum ## 🧭 Bib capsule Eric Provencher (@pvncher), "Choosing GPT-5.6 Sol, Terra, or Luna in Codex," X, 2026 — https://twitter.com/pvncher/status/2077708372363624894 ## 🧠 Author's thesis (Adler Q1+Q2) Provencher lays out OpenAI's official model-selection guidance for Codex's GPT-5.6 tiers — Sol (complex/ambiguous), Terra (pragmatic all-rounder), Luna (fast/well-scoped) — plus an orthogonal reasoning dial (Medium/High/xHigh/Ultra). The framework is a decision procedure: pick a model, then a reasoning level, then prompt it accordingly. Defaults matter — Sol Medium when unsure, and hold Ultra back for high-stakes/high-ambiguity planning rather than routine work. ## 🧩 Key arguments - Default to Sol Medium when unsure — a single unambiguous baseline before task-routing intuition develops. - Trade model size against reasoning depth: increase reasoning as models get smaller (Sol Medium ≈ Terra High ≈ Luna xHigh). - Reserve Ultra for stakes, ambiguity, or large context — explicit restraint against reflexive escalation to the most expensive tier. ## 🗒️ Notes — Literature atoms ### Default to Sol Medium when unsure which model to use > [!quote] > "If you're unsure where to start, go with Sol Medium." - **Where**: Opening line, positioned as the baseline before the per-model breakdown - **Why**: A single actionable default for anyone using Codex without a strong task-routing intuition yet. ### Increase reasoning level as model size decreases > [!quote] > "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." - **Where**: "Match the effort to the work" section - **Why**: A transferable heuristic for any multi-tier model system with an orthogonal reasoning dial — maps onto Claude's model+effort matrix. ### Most tasks don't need the top-tier Ultra mode > [!quote] > "Most tasks won't need Ultra... Save Ultra for when the stakes, ambiguity, or amount of context justify the extra depth." - **Where**: "What Ultra means" section - **Why**: Vendor restraint against reflexive escalation — matches the "delegate by default, escalate only when warranted" routing discipline. ## 🔖 Permanent candidates - Structure agent prompts as outcome specifications, not step lists — For capable agentic models on ambiguous work, specify goal, context, output boundaries, and finish-line criteria rather than a step-by-step itinerary, and let the model find its own path. (No existing Permanent covers the spec-vs-itinerary prompt structure — worth promoting; breaks down for narrow, well-scoped tasks.) ## Links - [[AI Cost & Tokenomics MOC]] — home MOC; the Sol/Terra/Luna + reasoning-dial framework is cost-vs-capability routing, the practitioner side of tokenomics - [[Fable 5 vs GPT-5.6]] — same model generation from the other vendor's framing; useful for cross-checking tier/effort guidance across ecosystems - [[Subscription Inference Is Sold Below Cost Creating a Tokenmaxxing Arbitrage]] — the "reserve Ultra" restraint is the user-side of the same cost pressure that makes flat-rate inference an arbitrage - [[Route AI Models by Temperament Not Just Cost Tier]] — the Sol/Terra/Luna reasoning-dial choice is temperament-plus-capability routing in practice, the principle stated in the Permanent. ## ✅ Processing checklist - [x] Claim atoms carried from deep-read analysis - [ ] Permanent candidates reviewed for promotion - [ ] Version History Diff check