← Help center

GPT models, reasoning effort and Fast: how to choose

Lesson 01 · Choosing a modelAbout 2 minutes
An illustrated introduction to GPT models, reasoning effort and Fast, in Chinese.
Chinese-language illustration from the original guide.

When choosing, consider task quality, waiting time and usage together. First distinguish three separate settings:

Model
Capability profile
Reasoning effort
Compute invested in the current task
Fast
Speed of the model service

01 / Choose a model for the task

Models, suitable tasks and starting reasoning effort
Recommended modelSuitable tasksStarting effort
GPT-6 LunaClassification, field extraction, format conversion and small code changesLow
GPT-6.1 SolEveryday development, multi-file edits and plans synthesized from source materialMedium
GPT-6.1 TerraPolishing writing, understanding images and charts, and operating software interfacesLow / Medium
GPT-6 AstraDifficult troubleshooting, architecture design, complex research and decision-makingMediumIncrease when needed
Swipe or scroll sideways to see the full table.

Based on OpenAI’s official model-selection guide. These combinations are practical starting points; available models depend on the interface you use.

02 / Match effort to the problem’s complexity

  • Low: Clear goals and few steps; prioritize speed.
  • Medium: Some planning is needed; balance speed and depth.
  • High / Extra High: Follow complex logic, compare options and check edge cases.
  • Max: A small number of particularly difficult individual tasks, when you can allow more time and resource use.

Higher effort usually uses more reasoning tokens and time. Judge any improvement in quality from the actual results. The same effort name does not mean the same capability across different models. Read the official explanation

In official Codex / ChatGPT Work, Ultra involves running sub-agents in parallel and suits large tasks that can be split up. Available levels vary by model and client. Max and Ultra

03 / Fast: trade higher usage for faster responses

When Fast is enabled in the current workspace, it keeps the selected model and reasoning effort and switches to an accelerated service tier. It suits frequent back-and-forth debugging and revisions. Actual costs depend on the interface you use; there is no universal fixed speed-up factor. Official Fast explanation

If most of the time is spent compiling, rendering or downloading, Fast offers limited benefit.

Compare quality and total elapsed time on the same task. Keep the lower-cost combination that meets your needs; increase effort or upgrade the model if requirements are missed, reasoning breaks down, or you keep having to redo work.