Prompting ChatGPT and OpenAI models
The GPT‑6 family offers the same large context window and core tool support at three distinct price tiers: Astra for maximum capability, Sol for code and agents, and Luna for efficient volume.
Choose the right member of the GPT‑6 family
| Model | API input / output | Choose it when |
|---|---|---|
| GPT‑6 Astra | $10 / $50 | the task is exceptionally difficult, long and valuable, and Sol does not pass the evaluation |
| GPT‑6 Sol | $2 / $10 | you are building coding agents, using tools or doing complex professional work |
| GPT‑6 Luna | $0.10 / $0.50 | the task is focused, the format is clear and volume, speed or price matters most |
All three publish a 1.05M token context window and up to 128,000 output tokens. Astra costs one hundred times Luna per text token, so difficult cases should be escalated rather than sending all traffic to it. One million input plus 250,000 output tokens costs about $22.5 on Astra, $4.5 on Sol and $0.225 on Luna at list price.
Remove repetition before adding more
Official model guidance recommends starting with something that works, removing one group of instructions or examples at a time, and running the same evaluations. State each rule once. Keep examples that encode a real product requirement or fix a measured failure.
Define what the model may do
For agentic tasks, distinguish answering, diagnosing and actually changing something. Name safe local actions and require confirmation before external, destructive, costly or scope-expanding actions.
Control reasoning and answer length separately
Test the default reasoning level and one lower level on the same tasks. Higher effort or pro mode belongs on work where the measured quality gain justifies the extra cost and time. Use response shape, required fields and API verbosity controls for visible length rather than requesting a long hidden reasoning trace.
Build for reuse and measurement
Put stable developer instructions, examples and tool schemas first and variable input later to benefit from prefix caching. Track cache reads, cache writes, total token use and accepted tasks together.
Sources
- OpenAI: Model guidance — GPT‑6 family, features and current prompting guidance
- OpenAI: Pricing — standard, cache, batch and long context
- OpenAI: Reasoning best practices — model-specific reasoning guidance
- OpenAI: Prompt caching — prefixes, cache boundaries and cost
- OpenAI: Evaluation best practices — representative tests