OpenAI maps GPT‑6 family for startups

The guide is framed as a decision aid rather than a launch note: OpenAI presents GPT‑6 as a suite with three named models, and asks teams to think of model choice and reasoning level as an intelligence/price tradeoff. That framing runs through the whole document — from the model list (Astra for the hardest reasoning, Sol for complex coding and research, Luna for repeated tasks like extracting invoice fields) to the reasoning levels (low, medium, high, and extra high/max) and speed options (Fast mode and Ultrafast, the latter available for GPT‑6 Astra).

Much of the operational advice concerns cost and context rather than raw capability. OpenAI points to prompt caching, saying cached input tokens cost up to 95% less than uncached ones depending on the model, and to compaction for reducing context size in longer conversations. It also tells teams to measure task success, latency, and cost per successful task before deploying, and to plan for monitoring and data controls.

For longer work, the guide leans on features that keep a run going rather than restarting it: mid-turn steering through the Responses WebSocket API, asynchronous tool calling, and multi-agent workflows in the Responses API via GPT‑6.1 Sol, which OpenAI says is currently in beta. Whether teams actually capture the caching savings and the long-running-task gains is likely to depend on how closely their prompts, skills, and repository instructions follow the guide — a discipline question as much as a model one.