A Model Guide for the GPT-6 Family: Astra, Sol, or Luna?

Hi, it's Shii-chan!

OpenAI News openai.com

With three GPT-6 models to pick from, it's easy to wonder which one to use. Today I'm sharing a guide that helps with exactly that!

What was announced?

OpenAI News published "A model guide for the GPT-6 family". It's aimed at startups and covers choosing GPT-6 models, tuning reasoning effort, improving prompts and skills, coordinating tools, and preparing workflows for production.

The guide has four main themes:

• Run effectively in production (manage context and cost with caching and compaction)
• Match the model to your workload (balance model, reasoning effort, and speed)
• Adjust your prompts and skills (be consistent about what to deliver and what counts as done)
• Keep long-running work on track (steering, async tools, delegation)

Why it matters

The GPT-6 family has several models, and the combination of model, reasoning effort, and speed changes both capability and cost a lot. Official guidance saves a lot of trial and error.

GPT-6 can also take on tasks that span hours or days, so updating instructions mid-run and not waiting on slow steps become important.

What changes

Instead of "use the strongest model on High for everything", you get a way to choose per task.

• Repeated work like extracting invoice fields can run on a lighter model with low effort
• Hard debugging or deep analysis is where you raise the model and effort
• Revisiting AGENTS.md and skills can cut down on needless questions and hesitation from agents

Dive Deep

Choosing a model

• GPT-6 Astra: the hardest reasoning work, where maximum intelligence is needed
• GPT-6.1 Sol: complex coding, research, and computer use
• GPT-6 Luna: focused tasks at scale with a clear goal, such as extracting invoice fields, classifying requests, or producing structured summaries

Reasoning effort

• Low: routine tasks like extracting facts or small edits
• Medium: work that needs judgment, like planning a feature or comparing options
• High: difficult debugging, deeper analysis, careful review
• Extra high / Max: try when High falls short, and keep only if the gain justifies the time and cost

For speed, Fast mode in the API gives faster, more consistent responses at a higher per-token cost than Standard. Ultrafast works in Codex and the API, speeds up token generation independently of reasoning effort, and is available for GPT-6 Astra.

Prompts and skills

• Keep skill descriptions short and explicit about when they run, load details only when needed, and replace rigid recipes with guidance
• In AGENTS.md, explain which documents and tests matter, and explicitly authorize safe routine workflows such as local tests with disposable data
• Replace blanket "always ask" rules with clear decision boundaries
• Define what "done" includes: implementing the change, running it, inspecting the result, and fixing failures

Long-running tasks

• Mid-turn steering: send corrections through the Responses WebSocket API while the model works. Updates are queued and don't cancel running tools or undo completed actions
• Asynchronous tool calling: the model keeps doing independent work while your app runs a slower task such as tests
• Delegation: GPT-6.1 Sol supports multi-agent workflows in the Responses API, assigning independent work to subagents (currently in beta)

In Codex, GPT-6 Astra can ask for clarification while it works, and you can steer an active task when requirements change.

Before production

• Cut context the task doesn't need and run independent tasks together
• Put stable instructions first and keep tool definitions consistent so caching works. Cached input tokens cost up to 95% less, depending on the model
• Use compaction for longer conversations
• Measure task success, latency, and cost per successful task on representative tasks

Computer use is available for GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna, and works even with apps that have no API. Prefer an API or connected tool when one can do the job directly.

Wrap-up

• Pick among Astra (hardest work), Sol (coding, research, computer use), and Luna (high-volume routine work)
• Start with Low to High reasoning effort, and use the top settings only when they pay off
• Write down what the agent may decide and what counts as done in your skills and AGENTS.md
• Use steering, async tools, and delegation to keep long tasks moving
• In production, rely on caching, compaction, and measuring success and cost

This one is especially for startups starting to build agents and apps on GPT-6!