What actually changed in GPT-6 Astra for documents, slides and spreadsheets?
GPT-6 Astra is OpenAI's frontier model released on 1 September 2026. For document work, the change is template adherence: Astra opens, edits and iterates on documents, spreadsheets and presentations while following a template you supply, matching its layout, tone and visual style rather than inventing its own.
That is a different job from writing well. Earlier models produced good prose and then poured it into a generic deck. Astra is trained to read your existing file as the specification.
OpenAI's own launch page calls it their best model for adhering to existing templates and for producing slides that convey key points with a structured narrative. The demo they published builds a deck about a fictional model using only a few slides from OpenAI's own presentation template, and holds the tone and layout throughout.
The numbers behind this are on AutomationBench, which tests real professional workflow tasks. Astra scores 41.4% against 18.1% for GPT-5.6 Sol. On OpenAI's internal design tasks it scores 50.0% against 47.4%, and on internal data science tasks 40.9% against 30.5%.
Availability: rolling out from launch day to a limited set of organisations, then to all ChatGPT Plus, Pro, Business and Enterprise users, plus the OpenAI API, Microsoft Azure and AWS Bedrock. API pricing is US$10 per million input tokens and US$50 per million output tokens. Enterprise administrators need to switch it on, because access is off by default at launch.
Why does giving Astra a template beat describing the format?
Describing a format in words forces the model to guess at spacing, hierarchy, colour and slide density. Handing it two or three real slides from your own deck removes the guessing. Astra reads the file, extracts the pattern, and applies it to new content.
This matters most for the things that make a deck look wrong at a glance. Font pairings, the position of your title block, how much text your organisation tolerates on a single slide, whether your charts sit left or right of the takeaway line.
The same logic applies to spreadsheets. Upload a workbook you already use, and Astra fills the structure you built rather than producing a fresh sheet with its own column order. OpenAI notes that Astra pulls only the context that matters into outputs instead of repeating information the work does not need.
Practical consequence for anyone working in Hong Kong on client-facing material: your brand template is now a reusable asset, not a formatting chore. One clean template file is worth more than a page of formatting instructions.
--- Attach 2 to 3 representative slides, not the whole 40-slide deck. Astra needs the pattern, not the archive.
--- Include one slide that shows your chart style, one that shows a text-heavy layout, and your title slide.
--- For spreadsheets, attach a filled example rather than an empty shell, so the model sees data types and formatting conventions.
How do you write a brief GPT-6 Astra can actually build from?
A vague brief now costs you more, not less. Because Astra is more capable of executing at length, a thin instruction produces a more elaborate wrong artefact. "Make me a deck about Q3" gets you twenty polished slides pointed in the wrong direction.
The brief is the highest-leverage input in the whole workflow. Five things belong in it: the audience, the decision you want them to make, the source material, the template, and the length ceiling.
Here is a complete brief you can paste and adapt. It is written for a presentation, and the same skeleton works for a report or a spreadsheet by swapping the artefact and the structure line.
Try this prompt
Build a 9-slide deck from the attached template and the attached Q3 performance data.
Audience: our regional marketing director, who has 10 minutes and has not seen the Q3 numbers.
Decision I want: approve or reject the proposed Q4 budget reallocation on the last slide.
Template: follow the attached 3 slides exactly for fonts, colours, title placement and chart style. Do not introduce new colours or icons.
Structure: slide 1 title, slide 2 the single headline number and what it means, slides 3 to 6 one channel each with one chart and one takeaway line, slide 7 what went wrong, slide 8 the reallocation proposal with figures, slide 9 the decision I am asking for.
Constraints: maximum 40 words of body text per slide. Every chart uses the actual figures from the attached file, no illustrative placeholders. If a figure is missing from the data, leave the field blank and list it on a final notes slide instead of estimating.
Before you build, tell me any assumption you had to make that would change the recommendation.
The last line is the part most people skip. It converts Astra's tendency to ask questions into a single, reviewable list instead of an interruption halfway through the build.
What does Astra do differently when your instructions are ambiguous?
Astra asks clarifying questions more often than GPT-5.6 Sol rather than assuming. OpenAI describes it as using context to fill routine gaps while asking focused questions when the answer could genuinely change the outcome. The trade-off is that it sometimes stops where you expected it to continue.
If you would rather it push forward, OpenAI's own model documentation supplies the fix. The published guidance tells you to instruct the model to infer intent and show a bias towards action, treating phrases like "can you", "I want to" or "help me" as calls to act rather than invitations to a follow-up question.
OpenAI's wording for this, from the GPT-6 Astra guidance published on 5 September 2026, is that the user should be approving a concrete, reviewable result. In other words: let it finish something, then review it.
Astra is also better at staying oriented when a task evolves. Earlier models often treated a mid-conversation steering message as a brand new goal and dropped the original constraints. Astra incorporates the new requirement, changes course when asked, and answers side questions without losing the broader task.
For a practitioner, that changes how you iterate. You no longer need to restate the whole brief every time you adjust one thing. "Make slide 6 a table instead of a bar chart" no longer risks resetting your 40-word text limit.
How do you stop Astra from writing like AI?
Astra defaults to lists, tables and Markdown formatting, and reuses the same phrases across sessions. If you want prose, you have to say so explicitly. OpenAI's published guidance goes further than most third-party prompt advice and includes an actual blocklist of what it calls slop words.
The named offenders in OpenAI's guidance include "delve into", "it's worth noting", "what's important is", "leverage", closing summaries like "In short:", and contrastive constructions of the form "X, not Y". It also flags invented hyphenated compounds and stock transitions.
This is unusually useful, because it is the model's own vendor telling you which habits to suppress. Paste this into a project instruction or a custom GPT and it applies to every output rather than one chat.
Try this prompt
Write in clear, concise paragraphs, each developing a single main idea. Use lists only when the information is genuinely parallel, sequential or easier to compare in a list. Avoid nested lists.
Use simple, familiar words, concrete examples and precise verbs. Favour active voice and direct statements. Make the main point early, then expand with the explanation and evidence the reader needs.
Do not use these phrases: "delve into", "it's worth noting", "what's important is", "leverage", "truly", "really", or closing summaries such as "In short" or "Conclusion:".
Do not use contrastive phrasing of the form "X, not Y". Do not invent hyphenated compound terms. Do not tell me what you are not doing or what you left unchanged. State the action directly.
Two caveats worth knowing. Astra also hands work to sub-agents less often than expected, so if you want parallel work you have to specify when and how much to delegate. And on coding tasks it over-tests, running suites out of proportion to small changes unless you tell it to rerun tests only when a new failure justifies it.
Where does this break down in real use?
Three failure modes are worth planning around before you rebuild a workflow on Astra.
Contradictory context beats your instructions
Astra follows long instructions better than its predecessors but is more sensitive to context. Unclear or contradictory instructions in the files it can read, including project instructions and skill files, can cause it to block work or veer off. OpenAI's recommendation is to audit every context document the model can access and give your direct instructions explicit priority.
OpenAI also publishes a debugging prompt for exactly this: ask the model to name the specific file and quote the specific instruction that caused it to pause or change direction. That turns an unexplained stall into a traceable cause.
Safety checks can pause legitimate work
Because Astra crossed the Critical threshold for cybersecurity capability under OpenAI's Preparedness Framework, misalignment monitoring runs in production for Astra-class models. OpenAI states plainly that extra safety checks can sometimes slow, pause or stop legitimate work. In ChatGPT you may be asked to review an action before it continues; in the API the task stops.
A large context window is not comprehension
Astra's long-context retrieval is strong, scoring 96.3% on OpenAI's MRCR v2 8-needle test in the 512K to 1M range. Retrieval is not the same as understanding a badly organised document. A messy 200-page source file still produces a confused deck. Clean the input.
One honest note on model versions: everything above reflects Astra at launch in early September 2026. OpenAI has said the Codex note-keeping behaviour becomes default in the coming weeks and that access expands over time, so re-test any workflow you build against the version you are actually served.
What should you try in the next 20 minutes?
Take a deck you have already delivered and are proud of. Delete everything except the title slide, one chart slide and one text-heavy slide. Save that three-slide file as your template.
Then take a piece of work you have not started yet, and run the brief from the third section above against it. Compare the result to what you would have produced by describing the format in words.
Do the same for a spreadsheet: keep one filled example, throw away the rest, and ask Astra to extend the structure with new data. The gap between "described the format" and "handed over the file" is where the whole capability lives.
If you want to see how the same principle plays out on the retrieval side rather than the output side, our piece on why AI gets worse the longer you chat covers what happens when the context you hand over is the wrong shape. And if you want a baseline read on where your own AI technique actually sits, the AI IQ Test takes a few minutes.
Template adherence is a small-sounding feature with a large practical consequence: the skill that now pays is briefing, not prompting tricks. We understand AI. We understand you better. With UD by your side, AI doesn't feel cold.
Reviewed by the UD AI team.
Turn One Good Prompt Into a Repeatable Workflow
Knowing the technique is step one. Making it run the same way every week, across your team, with your templates and your data, is the harder part. UD has spent 28 years helping Hong Kong teams put technology into daily practice, and we'll walk you through every step, from tool setup to workflow design and deployment.
