Last reviewed 29 August 2026. OpenAI's current developer guidance lists the GPT-5.6 family and recommends leaner, outcome-focused prompts. Your ChatGPT interface may use different labels or offer a different selection.
| Mode you see | Good for | How to prompt it | Watch out for |
|---|---|---|---|
| Quick or default | Emails, summaries, rewriting, everyday questions | Be direct. Include the task, essential facts and answer format. | Do not overload a simple job with a page of instructions. |
| Reasoning or thinking | Multi-step comparisons, planning, analysis and difficult problems | Define the outcome, constraints, evidence and what success looks like. Let the model choose the path. | Do not ask for hidden chain-of-thought. Ask for a concise explanation or checks instead. |
| Deep research or web-enabled | Current information, evidence reviews and questions needing several sources | Define acceptable sources, date range, citation rules and when enough research is enough. | Do not assume every citation is correct—open and check the important ones. |
| Creative or image-capable | Ideas, drafts, visual concepts and work involving an uploaded image | Name the audience, mood, style, constraints and a few reference qualities you like. | Do not rely on a single vague word such as “professional” to define the look or tone. |
Newer reasoning models often need a destination more than a script.
For a difficult task, explain the result you need, the boundaries, the evidence it can use and how you will judge success. Detailed step-by-step instructions are still appropriate when the exact process matters, but they can otherwise add noise.
Outcome: Compare these three options and recommend the best fit for my situation. Context: [facts about my situation] Success criteria: - use current, reliable evidence - explain the decisive trade-offs - identify anything uncertain - finish with one recommended next step
What “different versions” means in practice
The product can route work for you.
The default ChatGPT experience may select capabilities automatically. Start with a clear task before worrying about names.
Reasoning depth is separate from answer length.
You can ask for careful analysis and still request a short conclusion, a five-row table or three bullets.
Availability changes.
Plan limits and model labels are not permanent. We date this guide and link to official sources rather than promising access.
Official sources used for this guide
We base changing model claims on official OpenAI documentation and explain product uncertainty plainly.