Your everyday starting point: Choose Terra · Medium reasoning · Standard speed. For a simple, clearly defined task, choose Luna · Low. Most routine faculty and staff work does not need Sol, Astra, Max, or Ultra.
This is a practical starting recommendation for SBS employees using Standard ChatGPT Business seats, not a requirement to use the same settings for every job. Available model names and controls vary by product, seat, and workspace. The model, reasoning, and speed controls described here apply where they are offered in ChatGPT Work or Codex.
Match the model to the task
| Choose | When the task is… | Faculty or staff example |
|---|---|---|
| Luna | Clear, bounded, and easy to check. | Reformat an agenda, extract dates, sort a list into supplied categories. |
| Terra — everyday default | Ordinary writing, analysis, research, or practical file work. | Draft an announcement, compare approved procedures, analyze a small workbook. |
| Sol — deliberate escalation | Ambiguous or complex, with difficult tradeoffs or conflicting evidence. | Reconcile several inconsistent source documents after a focused first pass. |
| Astra — exceptional work | An unusually demanding task needing strong judgment across several steps and tools. | Investigate a difficult technical problem that simpler approaches have not resolved. |
Capability is not a reason to select the strongest model for everything. A short rewrite usually needs clear instructions more than additional intelligence. A consequential decision still needs the appropriate human review, whichever model you choose.
Check three settings before you send
- Model: return to Terra or Luna after completing an exceptional task. Do not assume a new conversation selected the settings you intended.
- Reasoning: start at Low for mechanical work or Medium for everyday analysis, where those choices exist. Raise effort when the task needs deeper planning or difficult reasoning.
- Speed: leave Standard selected unless reduced waiting time justifies higher consumption. Fast is a latency choice, not a quality upgrade or a way to save capacity.
- When a task contains subtle dependencies, challenging edge cases, or unresolved contradictions. First improve the sources and instructions; extra reasoning cannot supply missing evidence.
- For exceptional tasks with a clear reason for the additional work. Avoid them for routine emails, summaries, formatting, and simple file changes. Parallel work also adds usage and coordination.
- Check that you are in the intended SBS account and product surface. Availability depends on your seat, workspace controls, client, and rollout. Ask Technology Services if access differs from what you expect.
A practical escalation example
Task: turn an approved event agenda into a staff checklist.
- Start with Luna/Low: specify the columns and ask it to preserve dates and owners.
- If the result misses a requirement, give one precise correction.
- If the sources conflict and require interpretation, move that specific comparison to Terra.
- Use Sol only if meaningful complexity remains. A larger model is not the automatic next step after a weak prompt.
Use a focused prompt
Turn this approved agenda into a checklist for staff. Use columns: action, owner, due date, source section. Keep names and dates unchanged. Mark missing information as unknown. Produce one version. Check each row against the agenda.
Keep your usage predictable
Attach relevant files, set a clear output length, and request focused revisions. Work and Codex share the applicable usage allowance; a five-hour reset does not necessarily restore an exhausted weekly budget. Check your usage dashboard for the current allowance and reset times. Premium capacity is useful for sustained heavy work, but it does not make the strongest model necessary for ordinary tasks.
Related guides
Hit a five-hour or weekly limit? Follow the recovery checklist.
Write better prompts · Review and improve a response
Official guidance
OpenAI's models, usage, and speed documentation. Checked September 16, 2026. Use the choices and limits displayed in your account when labels differ.
Related to