OpenAI publishes guide for choosing and steering GPT-6 models

The guide lays out model selection, faster response options and ways to keep long tasks under control.

OpenAI has published a developer guide for the GPT‑6 family that explains how to choose between models such as GPT‑6 Astra, GPT‑6.1 Sol, and GPT‑6 Luna based on capability, latency, and cost, and how to tune reasoning level and speed modes for different workloads. The guide recommends using GPT‑6 Luna for high‑volume, repeatable tasks with clear outputs, and using Fast mode in the API when response time is more important than per‑token cost. It outlines prompt and agent design practices—like defining skills, decision boundaries, and “done” criteria, and documenting them in files such as AGENTS.md—to let models act more autonomously while keeping humans in control of approvals. For long‑running tasks that can stretch over hours or days, it details how to use asynchronous tool calls and GPT‑6.1 Sol’s beta multi‑agent workflows in the Responses API, as well as Astra’s ability in Codex to ask clarification questions and be steered mid‑task. The guide also highlights “computer use,” which allows GPT‑6 Astra, GPT‑6.1 Sol, and GPT‑6 Luna to operate web and desktop applications directly when no API is available, complementing conventional API and tool integrations.