OpenAI
Rewards restraint. The current generation performs better with shorter, less repetitive prompts than its predecessors did.
How to prompt it
Say each thing once. Remove repeated instructions and duplicated examples — OpenAI names this as a direct performance and efficiency gain.
Structure with Markdown headings. The model reads sectioned prompts reliably; you do not need XML here.
Set detail level with the verbosity setting where you have API access, and keep task-specific length rules in the prompt.
Describe tone as concrete choices — "state the answer first, then the caveat" — not as adjectives.
State autonomy boundaries once, in one place, rather than sprinkling approval requirements through the prompt.
Do not add "think step by step". Reasoning is already on, and the instruction mostly buys you tokens and latency.
Strengths
- Strong instruction following and structured output, including strict JSON schemas.
- Reasoning is built in and on by default; you rarely need to ask for step-by-step thinking.
- A response-level verbosity control, so you can set default detail outside the prompt text.
- Three tiers on one prompt format, so you can move between cost and capability without rewriting.
Weaknesses
- Repeating an instruction can make behaviour worse, not better — OpenAI documents that repeated "ask first" style rules cause unnecessary approval requests.
- Prompts carried over from GPT-4-era models are usually over-scaffolded and need trimming.
- Vague tone words ("professional", "friendly") land less reliably than concrete writing rules.
Models
GPT-5.6 Sol
Flagship for complex professional work.
- Context
- 1.05M
- Max output
- 128K
Good for: Hard analysis, long documents, agentic workflows with many tool calls.
Watch out: Highest cost and latency of the three. Reserve it for work where a marginal quality gain actually changes the outcome.
GPT-5.6 Terra
Balances intelligence and cost.
- Context
- 1.05M
- Max output
- 128K
Good for: The sensible default for most drafting, summarising and reasoning work.
Watch out: On genuinely hard multi-step problems the flagship still pulls ahead.
GPT-5.6 Luna
Built for budget-conscious, high-volume work.
- Context
- 1.05M
- Max output
- 128K
Good for: Classification, extraction, routing, and anything you run thousands of times.
Watch out: Give it tighter, more explicit prompts than you would give the larger tiers.
Sources and grounding
- 1Model guidanceOpenAI (Developer docs)Per-family prompting guidance, including lean prompts, autonomy boundaries and verbosity.
- 2ModelsOpenAI (Developer docs)Live model list, context windows and output limits.
- 3Prompt engineeringOpenAI (Developer docs)General prompting reference.