# Choosing a Codex Model and Reasoning Effort

Use the lightest model and effort that can complete the work reliably. A skill supplies the workflow and standards; the model and effort determine how much capability and reasoning are applied to that workflow.

## Rule of thumb

**Luna = Find / Generate / Repeat.**

Use **Terra with Medium reasoning** as the team default for approximately 70–80% of development work. Move to **Sol with High reasoning** when mistakes would be expensive or the work crosses complex system boundaries. Use **Max** only for exceptionally difficult, quality-first work after High is insufficient.

## Quick cheat sheet

| Model | Best fit | Typical reasoning |
|---|---|---|
| GPT-5.6 Terra | Normal C# or VB.NET work, Web Forms, Web API, SQL procedures and functions, ordinary debugging, Telerik or Kendo, unit tests and refactoring | Medium |
| GPT-5.6 Sol | Complex production bugs, architecture, large refactors, SQL performance, authentication or session problems, cross-project migrations, deep root-cause analysis and final review of important changes | High; Max only for exceptionally hard work |
| GPT-5.6 Luna | Search and exploration, finding usages or dependencies, boilerplate, simple SQL or CRUD, renames and transformations, documentation, basic tests and research tasks | Low to Medium |
| GPT-5.3 Codex Spark | Rapid interactive edits, tiny experiments and very small changes where instant feedback matters most | Low |

## Reasoning effort

| Effort | Use when | Examples |
|---|---|---|
| Low | The task is trivial, tightly scoped and easy to verify | Find a symbol, rename a field, reformat text, generate a small repetitive block |
| Medium | The task needs normal implementation judgment | Add a procedure, change a Web Forms page, build a form, produce a technical specification |
| High | The task has meaningful ambiguity, multiple layers or material regression risk | Diagnose a production defect, validate a migration, design a payroll integration, review a release package |
| Max | The task is exceptionally difficult and quality matters more than time or cost | A rare, deeply coupled failure after High has not produced sufficient confidence |

## How to use the recommendation on a skill card

1. Start with the model and effort shown on the skill card.
2. Lower the effort for a clearly repetitive or well-specified subset of the work.
3. Raise the effort when evidence is incomplete, the change crosses applications, or failure could affect production data or security.
4. Use Sol High for the final review of an important Terra implementation when independent deeper scrutiny is worthwhile.
5. Do not select Max automatically. Reserve it for the hardest quality-first work.

## Practical examples

- Use **Luna Low** to locate every reference to a setting or generate repetitive documentation.
- Use **Luna Medium** for a guided video script, email-template formatting or PDF document task.
- Use **Terra Medium** to build a rerunnable SQL package, Dynamic Form, API integration or normal Web Forms enhancement.
- Use **Sol High** for a platform merge, difficult performance problem, payroll interface, production root-cause investigation or final release validation.
- Use **GPT-5.3 Codex Spark Low** for a tiny interactive experiment, then move to the skill card recommendation if the change grows beyond that scope.

These recommendations are HRFocus team defaults. Adjust them when the task’s actual risk and complexity justify it.