A practical operating model for moving from the GPT-5.5 mental model to the GPT-5.6 family, with guidance for ChatGPT Work, Codex, model selection, reasoning depth, and explicit subagent orchestration.
The new model system has two decisions: choose the model family, then choose the reasoning effort. The model family controls the capability, speed, and cost profile. The reasoning effort controls how hard the selected model works.
Do not force a one-to-one conversion from GPT-5.5 Low, Medium, High, and Extra High. OpenAI states that there is no exact mapping from GPT-5.5 reasoning efforts to GPT-5.6. The right migration pattern is to test a familiar task at a lower setting and adjust upward when needed.
For most professional work, use Terra Medium as the default. For architecture, research synthesis, hard coding, and final review, use Sol High first. Move to Sol Extra High or Sol Max only when the work is genuinely difficult or high-value.
Use Luna for high-volume, bounded, repeatable work where the desired output is clear. It is best for extraction, summarization, classification, formatting, cleanup, routing, and structured transformations.
Best fit: fast operational work and worker-agent tasks.
Use Terra for everyday serious work where judgment matters but maximum depth is not needed. It is the natural starting point for business analysis, routine coding review, planning, research triage, and document development.
Best fit: GPT-5.5 Medium and many GPT-5.5 High use cases.
Use Sol when the work requires deep synthesis, high confidence, complex reasoning, broad context, or final judgment. This is the preferred model for architecture, cybersecurity, deep research, advanced coding, and executive-grade synthesis.
Best fit: GPT-5.5 High, Extra High, Max-style work, and final review.
This is a practical usage map, not an official numeric benchmark. It combines model family and reasoning depth to show where each setting normally belongs in day-to-day work.
| Old GPT-5.5 habit | New starting point | Use when | Adjustment rule |
|---|---|---|---|
| GPT-5.5 Instant or Low | Luna Low or Terra Low | Quick answers, summaries, formatting, simple transformations. | Use Luna when the task is repeatable. Use Terra when judgment matters. |
| GPT-5.5 Medium | Terra Medium | Standard professional work, planning, analysis, everyday coding, document review. | This is the safest new default for normal serious work. |
| GPT-5.5 High | Sol High or Terra High | Multi-step reasoning, tradeoffs, implementation planning, complex writing. | Use Sol when quality matters more than speed, latency, or usage. |
| GPT-5.5 Extra High | Sol High first, then Sol Extra High | Architecture, difficult debugging, deep synthesis, complex decisions. | Test Sol High first because GPT-5.6 may need less reasoning effort. |
| GPT-5.5 Pro Standard | Sol High or Sol Extra High | High-value analysis or review where reliability matters. | Use higher effort only when the decision quality justifies the cost. |
| GPT-5.5 Pro Extended | Sol Extra High, Sol Max, or Sol Ultra | Hardest tasks, high-stakes review, broad research, multi-domain project work. | Use Max for one hard problem. Use Ultra when the work can be split into lanes. |
Use for simple tasks, direct answers, cleanups, transformations, and predictable output formats.
Use for everyday professional work where you need useful judgment without deep analysis.
Use for multi-step reasoning, implementation planning, tradeoffs, coding review, and structured strategy.
Use when ambiguity, technical risk, architectural consequences, or executive quality matters.
Use when depth matters more than speed or usage, and the task is one coherent problem rather than parallel work.
Use when the task naturally decomposes into independent lanes that can be worked in parallel and synthesized.
| Mode | Subagents available? | How they start | Best use |
|---|---|---|---|
| Terra or Luna Low through High | Yes | You ask directly, or project and skill instructions request it. | Efficient scans, extraction, triage, test review, issue mapping. |
| Sol High or Extra High | Yes | Usually explicit unless Ultra is selected. | High-value reasoning with controlled delegation. |
| Sol Max | Yes | Explicit delegation, unless the environment supports proactive delegation separately. | Hard single-threaded work with occasional delegated evidence gathering. |
| Ultra | Yes | Can proactively delegate when parallel agents would materially improve speed or quality. | Broad reviews, multi-lane implementation planning, large project assessment. |
Start: Sol High
Escalate: Sol Extra High or Sol Max when the work requires deep tradeoff analysis.
Use Ultra: When strategy, finance, risk, technical feasibility, and execution planning can be split into separate lanes.
Start: Sol High
Worker agents: Terra High for codebase scans, test review, data migration review, and documentation review.
Use Ultra: For large branch review, modernization planning, and multi-discipline implementation assessment.
Start: Terra Medium
Escalate: Terra High or Sol High if the answer needs deeper judgment.
Use Luna: For cleanup, extraction, formatting, and high-volume processing.
Start: Sol High
Escalate: Sol Extra High or Sol Ultra for broad research with multiple independent lanes.
Worker agents: Terra Medium or Terra High for source triage and evidence extraction.
Review this branch with parallel subagents.
Spawn:
1. Architecture reviewer for architecture, boundaries, and migration risk.
2. Test reviewer for regression coverage, smoke tests, and fixture gaps.
3. Data reviewer for persistence, storage seams, and migration concerns.
Wait for all agents. Then synthesize the findings into a prioritized implementation plan with file references.Use the main agent as the final synthesizer.
Delegate bounded read-heavy work to Terra worker agents:
- One agent maps the codebase structure.
- One agent reviews the test strategy.
- One agent extracts data and persistence concerns.
Keep worker outputs concise. Return only evidence, risks, and recommendations.Move noisy exploration into subagents.
Do not flood the main thread with raw logs or intermediate notes.
Return a clean summary with:
1. What was inspected
2. What was found
3. What is risky
4. What should be done nextUse custom agents when you want reusable expert roles. The main idea is to keep the parent agent focused on synthesis and use worker agents for bounded evidence gathering, scanning, and review.
# .codex/config.toml
[agents]
max_threads = 6
max_depth = 1# .codex/agents/architecture-reviewer.toml
name = "architecture_reviewer"
description = "Reviews architecture, boundaries, state management, and migration risk."
model = "gpt-5.6-terra"
model_reasoning_effort = "high"
sandbox_mode = "read-only"
developer_instructions = '''
Review like a senior software architect.
Focus on architectural seams, domain boundaries, data persistence, regression risk, and migration path.
Return concrete findings with file references.
'''| Question | Use this choice |
|---|---|
| Is the task fast, structured, and repeatable? | Luna Low or Luna Medium |
| Is this normal professional work where judgment matters? | Terra Medium |
| Is this a serious analysis, plan, or coding review? | Terra High or Sol High |
| Is this architecture, executive synthesis, or high-risk judgment? | Sol High, then Sol Extra High if needed |
| Is this one hard problem that needs maximum depth? | Sol Max |
| Can the work be split into independent expert lanes? | Ultra, or explicit subagents on a non-Ultra setting |
| Do you want controlled orchestration instead of proactive delegation? | Explicit subagents through the prompt, AGENTS.md, skills, or custom agent files |
These notes preserve the official source basis used for this guide. OpenAI product details can change, so use these links as the first place to verify current availability, naming, and limits.