Choosing GPT-5.6 Sol, Terra, or Luna in Codex  🔗 **Source:** [Choosing GPT-5.6 Sol, Terra, or Luna in Codex](https://twitter.com/pvncher/status/2077708372363624894/?rw_tt_thread=True) > [!note]+ Document Notes > tags: [openai-codex, ai-model-selection, agentic-coding, prompt-engineering, subagent-orchestration] > author: Eric Provencher > source: X (Twitter) > reference: https://read.readwise.io/read/01kxnnng0mvs4e64kh708wrwp9 > date created: 2026-07-16 11:20 > > # 🏷️ Title: Choosing GPT-5.6 Sol, Terra, or Luna in Codex > <!-- deep-read:sections start --> ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ### 🔗 References - No external links included in the tweet thread itself; all guidance is self-contained product documentation. ## 📝 Summary - OpenAI's Codex now ships GPT-5.6 as three distinct model tiers — Sol, Terra, Luna — each tuned for a different work profile, plus a reasoning-level dial (Medium/High/xHigh/Ultra) layered on top of each. - **Sol**: complex, ambiguous, high-value work — deep investigation, cross-problem connection-making, hard debugging. **Terra**: the pragmatic all-rounder for everyday implementation and multi-step work with good judgment. **Luna**: fast, for clear well-scoped work — extraction, classification, transformation, structured summaries, and bounded implementation at xHigh reasoning. - **Sol Ultra** is the ceiling: maximum reasoning plus proactive multi-agent collaboration, explicitly positioned for planning (pulling together Slack threads, issues, PRs, docs, git history into an implementation plan) rather than routine work — "most tasks won't need Ultra." - Default recommendation for the undecided: Sol Medium. General rule of thumb: increase reasoning level as model size decreases (Sol Medium ≈ Terra High ≈ Luna xHigh in capability-per-task). - Subagents inherit the parent agent's model family and reasoning level by default; per-subagent customization (lighter settings for context-gathering, stronger for implementation) is announced as "coming later this week" at time of posting — not yet available. - Bottom line: this is OpenAI's official model-selection guidance for Codex power users, structured as a decision framework (which model → which reasoning level → how to prompt it) rather than a benchmark comparison. ## 📖 Main Content The post is structured as practical guidance for choosing among Codex's three GPT-5.6 model variants, aimed at users who already use Codex regularly and need a mental model for model/effort selection rather than a first-time onboarding. It opens with a moonshot/straight-shot framing (some tasks need deep planning and coordination, others are simple) and immediately gives a default: if unsure, start with Sol Medium. It then profiles each model by use case rather than by benchmark score — Sol for ambiguous/high-value work where "connecting ideas across a problem" and catching easy-to-miss details matters; Terra as the everyday implementation workhorse that "converges on a solid result without pushing for every last detail"; Luna as the speed option for high-volume, well-defined work, capable of "more substantial implementation work" when scope is clear, especially at xHigh reasoning. A dedicated section explains Ultra as a qualitatively different mode — not just "more reasoning" on Sol, but "proactive multi-agent collaboration" combined with maximum reasoning, explicitly pitched for planning tasks (synthesizing Slack threads, GitHub issues/PRs, docs, code, and git history into an implementation plan) rather than routine implementation. The post is explicit that Sol Medium can also produce strong planning results if subagents are used proactively — Ultra is positioned as an escalation for stakes/ambiguity/context volume, not a default upgrade. The prompting guidance section ("Give Codex a clear finish line") reframes prompt-writing around outcome-specification rather than step-by-step instruction: a good prompt supplies Goal, Context, Output and boundaries, and Finish line — with an example "Sol Ultra planning prompt" that explicitly instructs the model not to implement yet, only to plan and present a reviewable HTML page. The closing section ("Match the effort to the work") gives the reasoning-level heuristic explicitly: increase reasoning as models get smaller (Sol Medium ≈ Terra High ≈ Luna xHigh), and offers three named alternative combinations (Sol Ultra / Terra High / Luna xHigh) mapped to specific task shapes (high-stakes-and-scattered-context / well-scoped-with-complexity / well-scoped-and-speed-matters). One structural note buried mid-post: subagents currently inherit the parent's model and reasoning level by default, and per-subagent tuning (e.g., lighter models for context-gathering, stronger for implementation) is announced as forthcoming ("later this week") rather than shipped — a caveat readers should track since it directly affects the "match the effort to the work" advice once it lands. ### 🗒️ Notes - Real-world corroboration: a July 16 2026 YouTube video (Theo, t3.gg — "I need you to hear me out (it's REALLY good)") discusses running "56 Soul" inside Claude Code — auto-generated captions garbled "GPT-5.6 Sol" into "56 Soul"/"5.6 Soul" throughout that transcript. This post confirms the correct model name and resolves that video's own flagged uncertainty ("is 56/Soul a specific named product?"). - The four-part prompt structure (Goal / Context / Output and boundaries / Finish line) is a reusable framework independent of Codex specifically — comparable in shape to other agentic-coding prompt frameworks that separate "what" from "how." - The reasoning-level substitution heuristic (Sol Medium ≈ Terra High ≈ Luna xHigh) is presented as a rule of thumb, not a benchmarked equivalence — no data is cited for the claim. ## 📚 References - No external sources, citations, or links appear in the tweet thread body — this is OpenAI/Codex-team-authored first-party guidance with no secondary sourcing. ## 🗣️ Paraphrase - Codex's GPT-5.6 offers three model tiers matched to task shape rather than one model tuned by a single "power" dial: Sol for ambiguous/high-value work, Terra as the reliable default for everyday implementation, and Luna for fast, well-scoped, high-volume work — with Ultra as a separate, expensive escalation specifically suited to planning and synthesis tasks that draw on scattered context. - The practical advice is to under-provision by default (start at Sol Medium, save Ultra for genuinely high-stakes/ambiguous work) and to write prompts as outcome specifications (goal, context, boundaries, finish line) rather than step-by-step instructions, since Sol is designed to discover its own path through ambiguity. ## 📰 Newsletter OpenAI quietly shipped a model-selection cheat sheet for Codex power users: GPT-5.6 now comes in three flavors — Sol (deep, ambiguous work), Terra (the reliable everyday workhorse), and Luna (fast, well-scoped work) — plus Sol Ultra for genuinely high-stakes planning that needs to synthesize a Slack thread, a pile of GitHub issues, and your git history into a real implementation plan. The practical takeaway: default to Sol Medium if you're not sure, and remember reasoning level substitutes for model size (Sol Medium ≈ Terra High ≈ Luna xHigh). One thing to watch: per-subagent model/reasoning customization is "coming later this week" per this post — until then, subagents just inherit whatever the parent agent is running. Bonus: this post retroactively solves a mystery from a widely-watched Codex-vs-Claude-Code teardown video where auto-captions mangled "GPT-5.6 Sol" into "56 Soul" throughout. ## ✅ Action Items #### DevRel - Developer advocacy: this model-selection framework (task-shape → model → reasoning level → prompt structure) is a strong template for any DevRel content explaining a multi-tier AI product lineup — worth studying the four-part prompt structure (Goal/Context/Output-boundaries/Finish-line) as a reusable teaching pattern. - Developer experience: the "subagents inherit parent model by default, per-subagent tuning coming later this week" caveat is worth tracking — flag when that ships since it changes the advice in the "match effort to work" section. - Community management: not applicable to this source — no community/support-channel content. #### OSPO - Not applicable to this source — no open-source governance, licensing, or SBOM content. #### Community - Not applicable to this source — no maintainer/contributor/sponsor content; this is a vendor product-guidance post. #### WordPress - Core AI: directly relevant to any Core AI team evaluating multi-tier model selection UX for their own agentic tooling — Codex's three-named-tiers-plus-reasoning-dial approach is a concrete design pattern to study or contrast against. - CLI: the prompt-structure guidance (Goal/Context/Output-boundaries/Finish-line) is portable to any CLI-based agent tooling documentation, including WP-CLI-adjacent AI workflows. #### The WPCC - Not applicable to this source. ### FAIR - Not applicable to this source. ## 🧪 Hypotheses - H1: Tasks routed to Sol for ambiguous/high-value work produce measurably fewer "wrong turns" in debugging scenarios than the same tasks routed to Terra or Luna, at matched reasoning levels. - H2: The reasoning-level substitution heuristic (Sol Medium ≈ Terra High ≈ Luna xHigh) holds approximately true across task types, rather than being an oversimplified rule of thumb that breaks down for specific task categories. - H3: Prompts structured around the four-part framework (Goal/Context/Output-boundaries/Finish-line) produce higher first-pass acceptance rates than unstructured/step-by-step prompts, holding model and reasoning level constant. - H4: Once per-subagent model/reasoning customization ships, users who set lighter models for context-gathering subagents and stronger models for implementation subagents will see reduced token cost with no measurable quality loss compared to uniform-model subagent trees. ### 📊 Methodology - H1: A/B test matched debugging tasks across Sol, Terra, and Luna at equivalent reasoning levels; measure iteration count to correct fix and total token spend. - H2: Benchmark task suite run at each model/reasoning-level combination named in the post; compare output-quality scores to test whether the claimed rough equivalence holds. - H3: Controlled prompt-rewrite study — take a sample of existing unstructured prompts, rewrite them using the four-part framework, run both versions against the same model/task pairs, and compare acceptance/rework rates. - H4: Once subagent customization ships, compare token cost and defect rate between uniform-model and tiered-model subagent trees on matched multi-lane tasks. ### ❓ Questions - What specific benchmark or internal evaluation data underlies the "increase reasoning as models get smaller" substitution heuristic — is this measured or a design-team rule of thumb? - How does Codex decide when to "bring in subagents early" versus running single-threaded — is this purely user-prompted, or does the model have its own trigger heuristics? - What does the promised per-subagent model/reasoning customization interface actually look like once shipped — skill-level config, per-prompt flags, or something else? - Is there a cost/latency comparison published anywhere between Sol Ultra and the other tiers, given the post only says Ultra "uses significantly more tokens" without quantifying it? ### 🔮 Future Research - A follow-up post or benchmark once per-subagent model customization ships, to see whether it changes the "match effort to work" guidance given here. - Independent (non-OpenAI) benchmarking of Sol/Terra/Luna against each other and against competing agentic coding models (e.g., Claude Code's Opus/Sonnet/Haiku tiering) on matched task suites. - Tracking whether the "coming later this week" subagent customization feature actually ships on the stated timeline — a concrete, checkable claim from this post. ### 💡 Implications - **Practice**: Teams standardizing on Codex should encode this model-selection framework (or their own equivalent) into shared prompts/skills/documentation rather than leaving model choice to individual developer intuition per task. - **Practice**: The "give Codex a clear finish line" prompt structure is a directly adoptable template for any team writing internal agentic-coding prompt guidelines, regardless of which underlying model/harness they use. - **Cost/economics**: The Sol/Terra/Luna-plus-reasoning-level matrix is effectively a cost-control lever — defaulting to Sol Medium and reserving Ultra for genuinely high-stakes work is presented as the economically sensible default, relevant to any team tracking agentic-coding token spend. - **Education**: This kind of first-party "how to choose" post is a useful reference artifact for onboarding new Codex users, but should be paired with independent verification since all claims are vendor-sourced with no external benchmarks cited. ### 🧠 Learning Styles Assessment - Reading level estimate: Flesch-Kincaid Grade ≈ 10–11 (technical product-documentation prose, moderate sentence length, domain-specific proper nouns — Sol/Terra/Luna/Ultra — used as unexplained brand terms after first introduction). #### 👀 Visual learners - Present: one header image (a stylized graphic accompanying the tweet) but no diagrams, tables, or comparison charts in the body text itself. - Improve: a simple comparison table (Model | Best for | Reasoning levels | Example task) would let visual/scanning readers triage faster than the current prose-paragraph-per-model structure. #### 🔊 Aural learners - Present: none — text-only tweet thread with no audio/video companion. - Improve: a short audio walkthrough (30-60 seconds per model) would suit this learning style, though may be disproportionate effort for a short-format post. #### 🗣️ Verbal learners - Present: strong use of characterizing language per model ("pragmatic all-rounder," "moonshots and everything in between") that gives each tier a memorable identity. - Improve: a one-line glossary distinguishing "model" (Sol/Terra/Luna) from "reasoning level" (Medium/High/xHigh/Ultra) would help readers who conflate the two axes on first read. #### 🧩 Kinesthetic learners - Present: the example Sol Ultra planning prompt is a genuine copy-paste-able artifact readers can try immediately. - Improve: equivalent example prompts for Terra and Luna (not just Sol Ultra) would let kinesthetic learners practice the full decision framework, not just the highest tier. #### 🧠 Logical learners - Present: the "match effort to work" section gives an explicit rule of thumb (reasoning increases as model size decreases) and three named alternative combinations mapped to task shapes. - Improve: a decision-tree or flowchart formalizing "task shape → model → reasoning level" would suit logical learners better than the current prose walk-through. #### 🧑‍🤝‍🧑 Social learners - Present: none — this is one-directional vendor guidance with no discussion prompt or community angle. - Improve: a linked discussion thread or examples-sharing channel for how teams are actually using each tier would extend social learning value. #### 🧘‍♀️ Solitary learners - Present: the post is fully self-contained and can be read/applied independently. - Improve: a self-check ("which of your last 5 Codex tasks would you now route differently?") would help solitary learners apply the framework to their own recent work. #### 🌿 Naturalistic learners - Present: none — no ecosystem/sustainability framing; not a natural fit for this content. - Improve: not applicable to this source. ### 📊 Accessibility Assessment #### Screen reader semantics - Present: standard heading structure in the rendered markdown (H2-level section breaks); the accompanying image has visible alt text ("Choosing GPT-5.6 Sol, Terra, or Luna in Codex") matching the title. - Improve: none needed at this structural level. #### Keyboard navigation - Not applicable to this source — static text/image content, no interactive elements. #### Contrast and readability - Not assessable from markdown/text content alone; would require visual review of the header image and the live tweet's rendering. #### Text resizing and reflow - Not applicable to this source at the text-content level; governed by the platform (X/Twitter or Reader) rendering it. #### Media and captions - Present: one static header image; alt text present and descriptive (matches title, not just "image"). - Improve: none needed — this is a reasonable accessibility baseline for a single decorative/title image. #### Forms and interactions - Not applicable to this source. #### Standards reference - No WCAG 2.2 AA gaps identified from the available text content; the one image present has appropriate alt text. ### 🌍 DEIB Assessment #### Cultural breadth - Not applicable to this source — technical product documentation with no cultural/regional framing to assess. #### Inclusive language - Present: plain, jargon-explained-in-context language; no exclusionary phrasing. - Improve: model names (Sol/Terra/Luna) are introduced without an explicit "these are marketing names for model checkpoints X/Y/Z" framing — a minor clarity gap for readers expecting version numbers rather than names. #### Neurodiversity support - Present: consistent section-per-topic structure (headers, bold lead-ins per model) aids scannability and predictable navigation. - Improve: the comparison-table suggestion above (Visual learners) would also aid neurodivergent readers who benefit from structured/tabular information over prose. #### Representation - Not applicable to this source — no individuals or groups represented/discussed beyond the single author. #### Psychological safety - Present: neutral, instructional tone throughout; no content warnings needed. ### ✅ Logical Validity #### Strong reasoning - The post consistently ties each recommendation to a concrete task-shape rationale (e.g., Luna's speed suits "high-volume workflows like extraction, classification" specifically because those tasks are well-defined and repetitive) rather than asserting capability rankings without justification. - The Ultra section explicitly scopes when NOT to use the top tier ("most tasks won't need Ultra," "large plans do not always need Ultra"), which is a notable restraint for vendor guidance — it argues against reflexively reaching for the most expensive option, strengthening credibility. #### Fallacies - No significant fallacies identified — this is procedural/prescriptive guidance rather than argumentative content, and it doesn't make comparative claims against competing products that would need to be checked for unfair characterization. The one soft spot: the reasoning-substitution heuristic ("increase reasoning as models get smaller") is presented as settled advice without citing supporting evaluation data — not a fallacy per se, but an unsupported empirical claim worth flagging. #### Tone of voice - Confident, instructional, second-person ("you," "your") — typical vendor best-practices register. No hedging language, presented as settled internal guidance rather than provisional advice. ### 🛡️ DARVO Analysis - Not applicable to this source — this is first-party product guidance with no accusation, conflict, or defensive framing present. ## 🎯 Conclusion - Codex's GPT-5.6 model-selection guidance gives users a task-shape-first framework (Sol for ambiguous/high-value work, Terra for everyday implementation, Luna for fast well-scoped work, Ultra as a planning-focused escalation) paired with a reasoning-level dial and a four-part prompt-structuring template (Goal/Context/Output-boundaries/Finish-line). The strongest, most checkable claim is the explicit restraint around Ultra ("most tasks won't need Ultra") — a notably non-upsell-y piece of vendor guidance. The weakest-supported claim is the reasoning-substitution rule of thumb, presented without benchmark data. - Practical next steps for readers: (1) default new Codex tasks to Sol Medium unless there's a clear reason to route elsewhere; (2) restructure existing ad-hoc prompts using the Goal/Context/Output-boundaries/Finish-line template; (3) watch for the per-subagent model-customization feature ("coming later this week" at time of posting) and revisit subagent defaults once it ships; (4) if using Claude Code or another harness alongside Codex, cross-reference this framework against that harness's own model-tiering guidance for a consistent internal standard. ## 🔄 Refactor Appendix — Choosing GPT-5.6 Sol, Terra, or Luna in Codex ### 📚 Literature Atoms (H3) #### [H1] Default to Sol Medium when unsure which model to use > [!quote] ID:rw_codex_gpt56_h1 > "If you're unsure where to start, go with Sol Medium." **Context & use**: - Opening line of the post, positioned as the default/baseline recommendation before the per-model breakdown. - Directly actionable: a single, unambiguous default for anyone using Codex without a strong task-routing intuition yet. **Tags**: #xref **Highlight**: [view highlight](readwise://) <!-- atomicity: 1 idea ✓ --> #### [H2] Increase reasoning level as model size decreases > [!quote] ID:rw_codex_gpt56_h2 > "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." **Context & use**: - "Match the effort to the work" section — the core substitution heuristic for trading model size against reasoning-level depth. - Useful transferable principle for any multi-tier model system with an orthogonal reasoning-effort dial (directly comparable to Claude's model+effort matrix). **Tags**: #xref #aphorisms **Highlight**: [view highlight](readwise://) <!-- atomicity: 1 idea ✓ --> #### [H3] Most tasks don't need the top-tier Ultra mode > [!quote] ID:rw_codex_gpt56_h3 > "Most tasks won't need Ultra... Save Ultra for when the stakes, ambiguity, or amount of context justify the extra depth." **Context & use**: - "What Ultra means" section — explicit vendor restraint against reflexive escalation to the most expensive/capable tier. - Directly relevant to Courtney's credit-routing discipline: matches the existing "delegate by default, escalate only when warranted" pattern already documented for Claude/Codex routing. **Tags**: #xref #addendum **Highlight**: [view highlight](readwise://) <!-- atomicity: 1 idea ✓ --> ### 🧠 Permanent Candidates (H4) #### [P1] Structure agent prompts as outcome specifications, not step lists **Claim** (one sentence, Courtney's voice): When prompting a capable agentic model for ambiguous work, specify the goal, available context, output boundaries, and finish-line criteria rather than a step-by-step itinerary, and let the model discover its own path. **Evidence pointer**: [[#H1]] **Scope & conditions**: Holds for models/harnesses capable of autonomous context-discovery and multi-step planning (Sol-tier and equivalent); breaks down for narrow, well-scoped tasks where step-by-step instruction is actually faster and more reliable than open-ended exploration (Luna-tier work). **Links**: - [[Credit routing model]] — because extends (the Goal/Context/Output-boundaries/Finish-line structure is a more granular version of the existing "freeze the spec, then delegate" pattern) <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Resolved: the Theo (t3.gg) video's flagged uncertainty about "56 Soul"/"5.6 Soul" (auto-caption garble) is confirmed here as GPT-5.6 Sol — update that note's Fleeting Queue item [F1] to note this is resolved. #xref - [F2] Track whether Codex's promised per-subagent model/reasoning customization ("coming later this week" as of this post) actually ships, and whether it changes this framework's guidance. #articleseed ### → Things 3 (max 2 actions) - [ ] Cross-reference this Sol/Terra/Luna/Ultra framework against Claude Code's own model-tiering guidance for a consistent personal/team standard — agent-tooling #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Tweets/Choosing GPT-5.6 Sol, Terra, or Luna in Codex.md|Choosing GPT-5.6 Sol, Terra, or Luna in Codex by eric provencher]] ## Highlights - 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.** ([View Highlight](https://read.readwise.io/read/01kxnr5k0tg1cfccfnmw80t4k0)) - Note: Default recommendation when unsure which of the three GPT-5.6 model tiers to use. - Tags: atom/literature/h1 - 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. ([View Highlight](https://read.readwise.io/read/01kxnr5nmqzcw964be9eaaw2ek)) - Note: Vendor restraint against reflexive escalation to the most expensive tier — matches existing credit-routing discipline. - Tags: addendum, atom/literature/h3 - **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. ([View Highlight](https://read.readwise.io/read/01kxnr5mew8tzxbrqheav54b34)) - Note: Core substitution heuristic: reasoning level trades off against model size. - Tags: aphorisms, atom/literature/h2