How do you get the best results from GPT-6?

GPT-6 Astra, the first model of OpenAI's GPT-6 generation, was announced on September 3 and is rolling out to ChatGPT Plus, Pro, Business and Enterprise accounts and to the API. Most good prompting advice still applies to it. What changes is the balance: OpenAI's own developer guide describes a model that behaves more like a careful colleague than an eager assistant. It sticks closely to written instructions, fills small gaps from context, and stops to ask when an answer could change the result.

That is useful, and it also changes where a prompt should put its effort. You get the best results by spending your words on four things: the outcome, the authority you are giving it, the questions you want back, and the style of the answer.

How Astra behaves, and what to write about it

OpenAI's guide lists several tendencies and a prompting response for each. Here they are in plain terms.

The first row matters most for everyday use. Earlier models often guessed and pressed on. Astra is more likely to pause, which protects you from a wrong guess but can also leave a task half done while it waits. OpenAI's suggested fix is to have the model finish everything it can and ask for approval only once it has something concrete for you to review.

A worked prompt

Imagine you run a small bookkeeping practice and want Astra to prepare a client letter about a change in filing deadlines, using a spreadsheet of client names and a short memo you wrote.

Using the attached memo and client list, draft the letter we will send to each client about the new filing schedule.

What I want: one letter template plus a table of which clients need the "early filer" paragraph (those with a fiscal year ending in March).

You may decide on your own: wording, order of paragraphs, how to merge client names.
Ask me before: stating any date or amount that is not in the memo, or leaving any client out.

Do all the drafting first, then bring me one list of open questions at the end rather than stopping partway.

Style: plain, friendly paragraphs a non-accountant can follow. No headings, no bullet points in the letter itself. Under 250 words.

Each block answers a question Astra would otherwise raise. "You may decide" removes the routine questions. "Ask me before" lists the few that really matter, here invented dates or amounts, which you would always want flagged. "Do all the drafting first" keeps it from pausing on the first doubt. The style line overrides its habit of formatting everything as lists.

For builders using the API

If you call Astra through the API as gpt-6-astra, the guide adds a few practical points. The temperature, top_p and top_logprobs settings are no longer accepted and should be removed. Astra does not offer a reasoning effort of none; if you used none or minimal with an earlier model, OpenAI suggests starting at low and comparing results. Tool calling needs the Responses API.

The guide also stresses auditing skills (saved instruction packages) and project files such as AGENTS.md that the model can read. Because Astra takes written guidance seriously, a forgotten line like "always confirm before editing" in an old file can make it stop repeatedly. OpenAI recommends stating that the user's instructions take priority, and even asking the model to name the file and line that made it pause.

Price is part of getting good results too. Priced at $10 for each million tokens you send and $50 for each million it writes, Astra is a premium model, though OpenAI reports it often finishes tasks in fewer output tokens than earlier models. Measure on your own tasks before moving a busy workload over.

Which old prompting habits can you drop?

You can stop spelling out every small detail. With earlier models, many people wrote long prompts covering every possible case because the model would otherwise guess badly. Astra fills routine gaps from context and asks about the important ones, so a short list of the decisions you care about usually beats a page of exhaustive rules.

You can stop tuning randomness. Developers who adjusted temperature and top_p to control style must remove them for Astra. Ask for the tone, length and format in words instead.

You can stop relying on the model to pick a format for you, then fixing it afterward. Astra leans toward headings and lists by default, so say up front what you want.

And it is time to retire stale instructions. Saved custom instructions, project files and skills written for older models may contain rules you no longer mean. A model that follows them faithfully makes those leftovers visible, so read them again and delete what no longer applies.