ChatGPTReasoning modelsLevel 3: Workflows

How Do You Prompt ChatGPT's Reasoning Models?

State the goal, the constraints, the output format and what "done" means, then let the model work. Don't tell it to think step by step. OpenAI's best-practices page says that may not help and can sometimes hurt, because reasoning models already reason internally.

Prompt to copy
Goal: [the outcome you need, in one sentence]

Context:
[paste or attach the material it needs]

Constraints:
- [hard limit, such as budget, deadline or tools allowed]
- [what must not change]

Output: [format and length, such as a one-page memo with a table]

Done means: [how you'll judge the answer]

Before you finish, check your answer against each constraint, and tell me which ones you couldn't confirm.
Do
  • State a clear goal
  • Give constraints and the output format
  • Say what done means and how to verify
Don't
  • "Think step by step"
  • "Explain your reasoning"
  • Spelling out every intermediate step
Based on OpenAI's reasoning guides.
  1. InstantGPT-6 Sol
  2. MediumIncluded on Plus
  3. HighIncluded on Plus
  4. Extra HighNeeds Pro, Business or Enterprise
  5. ProGPT-6 Pro, on the same plans
The Thinking slider in Chat, lowest to highest. Source: OpenAI's Help Center article on GPT-6, checked 6 November 2026.

Try it now

  1. 1Pick one hard problem from this week and fill in the prompt above.
  2. 2In Chat, select GPT-6 in the model menu and move the Thinking slider to High. In Work or the desktop app, use the control below the composer. Choose the Smarter preset, or open Advanced to choose a model and effort.
  3. 3Send the prompt, and answer any question it asks rather than restarting.
  4. 4It worked if the answer meets your "done" line and names any constraint it couldn't confirm.

Use it when

Problems with several steps where an early wrong assumption ruins the rest: planning a migration, comparing options with conflicting evidence, finding the cause of a bug, or checking whether a set of numbers holds together.

Skip it when

Rewriting a paragraph politely or pulling a date out of an email. A fast model does this well, and on Plus, OpenAI estimates only 5 to 45 messages per five hours on Astra in Codex and Work, compared with 350 to 3,000 on Luna. Skip it if you can't yet say what a good answer looks like, too. Sort that out first in a quick chat.

Why it works

OpenAI's docs say reasoning models use internal reasoning tokens before they respond, which helps them plan and handle ambiguity. The steps are the part they're good at, so your job is to set the target. OpenAI's reasoning guide recommends a clear goal, constraints and an explicit format without prescribing every intermediate step. For research-style tasks, it says to define what "done" means and how the model should verify its work.

Effort is a separate control, and it isn't a cure for a weak prompt. OpenAI's guide calls effort "a tuning knob, not the primary way to recover quality." In the ChatGPT desktop app and Work, the labels are Light, Medium, High and Extra High, with Max and Ultra on some setups. OpenAI suggests starting at High for Luna, Light for Astra, and the app's default for Sol.

Example

Say you need an on-call rotation for six engineers. Two can't do weekends, nobody should work back-to-back weeks, and it must cover 13 weeks. You state those constraints, ask for a table, and define done as "every week covered and every constraint met." You don't explain how to build the schedule. If the constraints can't all be met, the final check should say which one fails.

Going further

Chat uses different labels. OpenAI's Help Center lists a Thinking slider with Instant, Medium, High and Extra High, all using GPT-6 Sol, plus Pro, which uses GPT-6 Pro, powered by Astra. Plus includes Medium and High, while Extra High and Pro need Pro, Business or Enterprise. Free and Go choose Think for harder questions. On Plus, Astra is only in Work and Codex. The pricing page calls the two modes GPT Instant and GPT Reasoning. On the API, reasoning tokens stay hidden but use context and are billed as output, and OpenAI suggests reserving at least 25,000 tokens at first. Hit max_output_tokens and a response can come back incomplete with no visible text. Mode and effort are also separate settings: pro mode does more model work at higher cost.

Common mistakes

  • Raising effort to fix a vague prompt. Rewrite the prompt first, since OpenAI says effort isn't the main way to recover quality.
  • Adding "think step by step" out of habit from older models.
  • Leaving out the format. OpenAI's prompt guidance says GPT-6 Astra tends toward detailed, formatted responses, so state the length you want.
  • Pasting examples that don't closely match your instructions. OpenAI warns that mismatched examples can produce poor results.

Sources

Checked 6 November 2026. Products change often, so confirm details with the vendor.

Related

Beyond the course