Source: Reasoning best practices (OpenAI)

OpenAI's guide to when to use reasoning models and how prompting them differs: simpler prompts, no CoT instructions, zero-shot first.

Contents
  • What: official API documentation page
  • Publisher: OpenAI
  • Date: accessed 2026-11-14 (examples reference o1, o3, o3-mini, o4-mini, GPT-4o)
  • Original: Reasoning best practices

Summary

Positions reasoning models as “planners” for ambiguous, multi-step, document-heavy or evaluation tasks, and GPT models as fast, cheap “workhorses” for well-defined tasks, with a common pattern of planning on one and executing on the other. It then lists how prompting changes for reasoning models.

Key claims

  • From o1-2024-12-17, reasoning models take developer messages in place of system messages.
  • Keep prompts brief and direct. Asking reasoning models to “think step by step” is unnecessary and can hurt.
  • Use delimiters (Markdown, XML, section titles) to separate input parts.
  • Try zero-shot first; add few-shot examples only if needed, and make them match the instructions closely.
  • State explicit constraints and define what success looks like.
  • Reasoning models avoid markdown by default; a first line of “Formatting re-enabled” in the developer message turns it back on.

Feeds

Chain-of-thought · Few-shot prompting · System prompts

Suggest changes →