OpenAI publishes a practical guide to the GPT-6 family

OpenAI published A model guide for the GPT-6 family on 2 October 2026. The page is a product guide, not a new model release. Its own summary says to match the model to the workload, keep prompts and skills consistent, use caching and compaction for cost, and use steering, async tools, and delegation on long jobs.

Which model the guide points at

The guide says GPT-6 Luna is for focused, repeated work with a clear goal, and gives extracting invoice fields, classifying requests, and structured summaries as examples. It says API Fast mode is for cases where response time matters, such as chat or coding tools, at a higher per-token price than Standard. It says Ultrafast, available for GPT-6 Astra, speeds token generation independently of reasoning effort when the premium is worth it, in Codex and in the API. It does not reprint the price card. Prices remain the ones on OpenAI’s pricing pages, which this guide only points toward.

On cost, the guide says cached input tokens cost up to 95% less than uncached input tokens, depending on the model, and that stable instructions and tool definitions should sit before the changing task. It says compaction reduces context on long conversations while keeping the state needed to continue, and that a production check should measure task success, latency, and cost per successful task. Those are OpenAI’s instructions, tied to docs the guide links.

Long jobs and computer use

The guide says mid-turn steering on the Responses WebSocket API queues a correction without cancelling a running tool or undoing a finished action. Asynchronous tool calling lets the model keep doing independent work while the app runs something slower, such as tests, and the guide says to wait for that result before starting work that depends on it. It says GPT-6.1 Sol can assign independent subtasks to subagents in the Responses API and combine the findings, and that multi-agent is in beta. In Codex, it says GPT-6 Astra can ask a clarifying question during a run, and that you can say which work may continue while you are away.

Computer use, the guide says, lets GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna operate websites and desktop apps that have no API. For builders adding that to their own app, it names Playwright for browsers and PyAutoGUI for desktop apps. The guide names those libraries as options for builders. It does not say OpenAI tested either integration.

Customer stories the guide repeats

The page closes with four short customer notes and labels them as how teams are building with GPT-6 Astra. It quotes Harvey cofounder Gabe Pereyra on giving the model more context. It says Cognition uses Astra inside Devin and, in an iPhone-game example, returned a simulator recording plus a report of what was checked. It says Hex turns a sales question into findings and a dashboard. It says Invideo reports roughly three times the success rate on color-grading tasks, and that a few editors created about 50 effects in one day. Those figures are the customers’ or OpenAI’s retelling. The guide does not show the measurement method.

Practical takeaway

Use the page as a map of features OpenAI already documents: caching, compaction, Fast and Ultrafast, steering, and a beta multi-agent path on GPT-6.1 Sol. It does not replace the model cards or the price list. Treat the 95% caching line as “up to”, and the Invideo and effect-count lines as reported outcomes, then measure cost per finished task on your own prompts before you lock a workflow to one variant.