A model guide for the GPT-6 family

OpenAI has published a comprehensive model guide for the GPT-6 family of models, designed to help startups select appropriate models, manage reasoning effort, improve prompts, coordinate tools, and prepare workflows for production environments. The suite features GPT-6 Astra for maximum intelligence and difficult reasoning, GPT-6.1 Sol for complex coding, research, and computer use, and GPT-6 Luna for focused tasks at scale and everyday repeated work. Developers can adjust reasoning effort in the API across tiers ranging from low for routine tasks up to extra high for difficult debugging. To run effectively in production, the guide emphasizes using prompt caching, which can reduce cached input token costs by up to 95%, and compaction to manage context size in longer conversations. For long-running tasks spanning hours or days, the guide highlights features like mid-turn steering via the Responses WebSocket API, asynchronous tool calling, and multi-agent delegation supported by GPT-6.1 Sol. Additionally, the guide outlines best practices for establishing clear decision boundaries, updating instructions, and defining success criteria before deployment.