Best default for most people: Use Terra with Medium reasoning at Standard speed. Choose Luna for clear, routine work. Move to Sol only when the task is complex, ambiguous, or has resisted a simpler model.
Model, reasoning effort, and speed are three separate choices. Each can affect usage, so selecting the most powerful option for every request can consume limits faster without improving routine work.
The three choices at a glance
| Choice | Start here when… | Examples |
|---|---|---|
| Luna | The request is clear, repeatable, and easy to check. | Reformat text, extract fields, classify items, rename content, create a first-pass summary. |
| Terra | You need dependable everyday analysis or creation. | Draft a guide, analyze a spreadsheet, research a defined question, make routine code changes. |
| Sol | The work is ambiguous, high-stakes, integration-heavy, or unusually difficult. | Design architecture, solve subtle bugs, reconcile conflicting evidence, perform a critical final review. |
Choose reasoning effort
- Use for simple extraction, formatting, classification, direct questions, and small edits with clear acceptance criteria.
- Use for most writing, analysis, research, file work, and routine technical tasks. Medium usually provides the best balance of quality, time, and usage.
- Use when the task has subtle tradeoffs, important edge cases, difficult debugging, or consequential decisions. Give the model a specific goal and review criteria before increasing effort.
- Max is intended for a single unusually hard, quality-first problem. Ultra can coordinate parallel workstreams and can use substantially more resources. Most staff and faculty tasks do not need either setting.
Choose speed
Use Standard unless waiting time is genuinely blocking you. Fast mode can return results sooner, but eligible GPT-5.6 and GPT-5.5 work consumes credits at a higher rate. OpenAI currently documents Fast as approximately 1.5× the speed and 2.5× standard credits for those models; this can change.
Do not use Fast as a quality setting. It changes latency and usage, not the task’s difficulty. Increase reasoning only when the problem needs more thinking.
A usage-conscious escalation pattern
- Write a clear prompt with the outcome, useful context, expected format, and boundaries.
- Start with Luna/Low for mechanical work or Terra/Medium for normal knowledge work.
- Review the result and give one targeted correction.
- Move up one level only if the task—not merely the prompt—requires it.
- Use Sol/High, Max, or Ultra for exceptional difficulty or a deliberate final review.
How to make weekly limits last
- Describe the desired output and review criteria in the first prompt.
- Attach only the files and sources that matter.
- Start a fresh chat when an old conversation contains irrelevant context.
- Ask for a focused output instead of several optional versions.
- Use Standard speed unless the time savings justify the higher usage.
- Reserve Sol and higher reasoning for tasks where they materially improve the result.
About limits: Exact limits depend on the SBS plan, current OpenAI policies, task complexity, context size, tools, and whether work runs locally or in the cloud. Treat any numeric estimate as temporary and check the usage indicator in the product.
Related guides
Official guidance
See OpenAI’s current model guidance, speed documentation, and usage and pricing documentation.
Related to