Skip to content

OpenAI

OpenAI’s GPT-6 guide says drop blanket “always ask” rules

5 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

All about: OpenAI models and products →
Image: openai.com

OpenAI’s guide to building with the GPT-6 family, published October 2, asks teams to replace blanket “always ask” rules with stated limits on what an agent may do on its own and what needs approval. Teams whose prompts or AGENTS.md still make the agent check in about everything have rewriting to do. The guide pairs those limits with a definition of “done” that covers running the change, inspecting the result and fixing failures, for tasks it says can now span hours or days.

What to change in prompts and AGENTS.md

“Give the model a clear assignment,” says Eric Provencher of Developer Experience at OpenAI. The guide breaks that down into the result you want, who it’s for, the relevant context and constraints, and what counts as done. The decision rule comes with an example: the model can choose how to organize a summary but should check with you before changing the project’s scope. For AGENTS.md, the instruction file a coding agent reads from the repository, the guide asks teams to say when particular documents and tests are relevant and to authorize safe routine work explicitly, such as running local tests with disposable data and no production access.

Skills get the same treatment: short descriptions that say when each one should run, supporting detail loaded only when needed, and guidance in place of rigid recipes. The guide also wants a short handoff at the end of a task, covering what changed, what was checked and what still needs attention.

Picking a model and an effort level

The guide splits the family by job. GPT-6 Astra is for the hardest reasoning, GPT-6.1 Sol for complex coding, research and computer use, and GPT-6 Luna for focused, repeated work with a clear goal, such as extracting invoice fields, classifying requests or producing structured summaries. At launch, OpenAI priced GPT-6 Sol at $2 per million input tokens and $10 per million output tokens, and Luna at $0.10 and $0.50, half of their GPT-5.6 promotional pricing. A high-volume classification job moved from Sol to Luna pays a twentieth of the rate on both input and output.

Reasoning effort is the second dial in the API. Low fits routine work like extracting facts or making small edits, Medium fits work that needs judgment such as planning a feature, and High fits difficult debugging, deeper analysis or careful review. Extra high and Max are for testing where supported when High falls short, and the guide says to keep them only if the improvement justifies the added time and cost. In Codex, the advice is to start from each model’s default level, then lower it for simpler tasks or raise it for deeper analysis.

In OpenAI’s launch figures for Sol and Luna, GPT-6 Sol at xhigh effort scored 33.2% on AutomationBench at $0.27 per task, while GPT-6 Astra at low effort scored 30.3% at 3.9 times Sol’s cost.

Speed is a separate choice. Fast mode in the API gives faster, more consistent response times at a higher per-token cost than Standard processing, and the guide points it at chat apps and coding tools. Ultrafast, in Codex and the API, speeds up token generation independently of reasoning effort and is available for GPT-6 Astra.

Runs that last hours or days

In the API, mid-turn steering sends a correction through the Responses WebSocket API while the model is working. Corrections are queued and don’t cancel running tools or undo completed actions. Asynchronous tool calling lets the model carry on with independent work while your app runs something slow, such as tests, and returns the result when it’s ready; the guide’s condition is to wait for that result before starting work that depends on it.

Delegation to subagents is still in beta. It runs on GPT-6.1 Sol in the Responses API, which can hand independent work to subagents, for example investigating different parts of a codebase, and combine their findings into a final response.

Codex gets its own version of the same controls. With GPT-6 Astra, Codex can ask for clarification while it works, and the guide suggests resolving the questions that affect the next step while naming which independent work can continue. If you’ll be away, tell Codex which tasks can go on and when it should pause for your answer. When requirements change, steering the active task should say both what changes and what stays the same.

Computer use works with all three models on websites and desktop apps, including applications without an API. The rule is to use an API or connected tool when it can do the job directly, and computer use when the model has to read a screen, click buttons or fill in a form. Apps that build it in need to give the model a tool that runs code to control a browser or desktop: Playwright for browsers, PyAutoGUI for desktop apps.

Where the cost goes

Cached input tokens cost up to 95% less than uncached ones, depending on the model. When OpenAI’s GPT-6 caching update arrived in September, the figure was up to 90% off cached input. Caching reuses work only for the opening stretch of a prompt that exactly matches an earlier request, which is why the guide says to put stable instructions and reference material ahead of the details that change, and to keep tool definitions consistent. A caching dashboard and a diagnostics guide show where that reuse breaks down.

Cost estimates for a complete workflow should include cache writes and any long-context rates as well. For longer conversations, compaction shrinks the context while keeping the state needed to continue, and the guide also suggests cutting context the task doesn’t need and running independent tasks together so one slow step doesn’t hold up unrelated work.

Before deploying, the guide says to run representative tasks and measure task success, latency and cost per successful task, decide how you’ll monitor behavior, and review the data controls for your application, with OpenAI’s API deployment checklist as the next stop.

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.