# Can ChatGPT Astra Create Good Slides? An Honest Test for Professional Teams > GPT-6 Astra is the strongest model to date and the first OpenAI flagship that names slides as a headline capability. It still lacks the professional depth that consulting and finance decks need. Here is where, and how to test it. - Author: [Florian Ploszczyk](https://www.offgen.ai/en/authors/florian-ploszczyk) - Published: 2026-09-11 - Updated: 2026-09-11 - Category: PowerPoint & Agent Workflows - Labels: PowerPoint & Agent Workflows, Foundations, Consulting - Canonical URL: https://www.offgen.ai/en/blog/can-chatgpt-astra-create-good-slides ## Evidence for this article This article supports its claims with 9 sources. Key sources include: 1. [GPT-6 Astra: A new generation of intelligence](https://openai.com/index/gpt-6-astra/) (OpenAI) 2. [GPT-6 Astra: The next generation in intelligence for work](https://openai.com/index/gpt-6-astra-next-generation-work/) (OpenAI) 3. [GPT-6 Astra System Card](https://deploymentsafety.openai.com/gpt-6-astra) (OpenAI Deployment Safety Hub) [Full source list](#sources) Yes, GPT-6 Astra creates good slides. It is the strongest model I have used, and it is the first OpenAI flagship whose launch material names slides as a headline capability. For a storyline draft, a restructure, or a summary of source material, it is excellent. What it cannot do yet is produce the kind of deck a consulting partner or a deal team would sign. Its research only reaches the public web. OpenAI itself flags template fidelity as limited. Nothing on the slide tells a reviewer where a number came from, and nothing tells the model which pages, claims, or data it may touch. Those problems live in the workflow around the model, and a smarter model on its own leaves them where they are. This is my assessment as of 11 September 2026, one week after release. Tool evaluations age fast, so treat the dates as part of the claim. ## What GPT-6 Astra is [#what-gpt-6-astra-is] OpenAI released GPT-6 Astra to a first group of organisations on 3 September 2026 and to Plus, Pro, Business, and Enterprise users a day later. The API followed through OpenAI, Microsoft Azure, and AWS. OpenAI calls it [state of the art on computer use, browsing, software engineering, cybersecurity, science, and professional work](https://openai.com/index/gpt-6-astra/). It is also the first model OpenAI has rated as critical for cybersecurity under its preparedness framework, so some prompts get refused. The sentence that matters for this article sits in the [work announcement](https://openai.com/index/gpt-6-astra-next-generation-work/): OpenAI says Astra is its best model so far at sticking to existing templates and producing well laid out slides with a structured narrative. Earlier launches mentioned documents somewhere in the middle of the page. This one puts slides near the top. ## What OpenAI claims, and what independent testing found [#what-openai-claims-and-what-independent-testing-found] | Question | OpenAI position | Independent evidence, September 2026 | | ------------------------------------------- | --------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------- | | Is it the most capable model available? | "The world's most intelligent and aligned model" | [Artificial Analysis](https://artificialanalysis.ai/articles/benchmarking-gpt-6-astra) ranks it joint first on its Intelligence Index | | Is it better at long professional tasks? | Yes, including polished documents, spreadsheets, and presentations | Around 90 Elo points above GPT-5.6 Sol on AA-Briefcase, the long-horizon knowledge work evaluation | | Does it produce better presentation output? | Best model yet for template adherence and layout | Presentation Quality Elo declined versus GPT-5.6 Sol, which still leads all models on that sub-score | | Does it hallucinate less? | "Substantially fewer factual errors" according to the [system card](https://deploymentsafety.openai.com/gpt-6-astra) | Hallucination rate on AA-Omniscience roughly halved at maximum effort, and remains above zero | | Does it operate inside PowerPoint? | Yes, through the [ChatGPT for PowerPoint](https://help.openai.com/en/articles/20001242-chatgpt-for-powerpoint) add-in | In beta since May 2026; template and font handling and advanced formatting edits are documented as limited | Read the third row twice. In the same evaluation, Astra got better at the analysis and worse at the presentation. And the way it reaches PowerPoint is a beta add-in whose own help page says template handling is not fully there yet. I do not think either point takes much away from the release. They tell you where the model is ready and where the workflow around it still has to carry the load. ## Where Astra is already good enough [#where-astra-is-already-good-enough] Most decks are team updates, training material, status reviews, a first draft of a workshop agenda, a slide with too many words on it, or an old deck that needs a new order. ChatGPT and the PowerPoint add-in handle all of that well, inside the application, on every plan including the free one. If that describes most of your week, use it and take the time back. The rest of this article is about the smaller set of decks where the bar is higher. Even on those decks, Astra does several things better than any model before it. It builds a storyline. Give it a messy source pack and ask for an executive summary with supporting sections, and the result holds a [pyramid structure](/en/blog/pyramid-principle-powerpoint-consultants) from the first slide to the last. OpenAI says the model pulls only the context that matters into the output instead of padding it. That matches what I see. It copies a style from an example. Three slides from an existing deck are enough for Astra to carry the tone and density into a new deck on a different subject. That is what OpenAI means by template adherence, and it works. It invents fewer numbers. The lower hallucination rate sounds abstract until you think about the reviewer who reads the draft at the end of a long day. Fewer wrong figures in the first version means fewer wrong figures that get through. It works around the deck, not only inside it. Astra browses, reads attachments, runs analysis, and revises its own output. For a research-heavy first pass on a public topic, an analyst's afternoon becomes a few minutes. So against a normal internal first draft, Astra wins clearly, and for many teams that is the whole question. A client proposal, a board pack, or a regulated disclosure asks something different, and that is where the next four sections go. ## Where professional depth is still missing [#where-professional-depth-is-still-missing] ### 1. Research stops at the public web [#1-research-stops-at-the-public-web] Astra can only research what it can reach: the open web, the files you upload, and whichever connectors your workspace allows. A professional deck rests on material that never appears on the open web, such as licensed market data, project references, the pricing logic your partners use, and the version of last quarter's numbers that finance signed off. If that knowledge sits in folders, it is not in the prompt, and the model cannot reason about what it has not seen. The [Company Brain guide](/en/blog/powerpoint-company-brain-guide) goes into why retrieval over approved, described content is what turns an impressive draft into a usable one. There is a second problem with research, and it is about the trail rather than the reach. Astra cites what it browses, but the citation stays in the chat. It does not land on the slide as a source footer next to the claim. A reviewer still has to rebuild the trail by hand, and the [three checks for every generated citation](/en/blog/powerpoint-citations-source-footers) still apply: does the source exist, does it say this, and is it current. ### 2. Better analysis, weaker slides [#2-better-analysis-weaker-slides] I think the Artificial Analysis result has been underread. On the same long-horizon evaluation, Astra's analytical quality rose sharply while its presentation quality dropped below the previous model. My reading is that the model got better at deciding what belongs on a slide and slightly worse at making the slide. For professional teams that is the better way round. A well argued slide with uneven spacing takes seconds to fix. A polished slide with a wrong recommendation does not. But it means "Astra makes better slides" is only half true, and the half that is not true is the half you notice first in a demo. ### 3. The add-in does not guarantee your master [#3-the-add-in-does-not-guarantee-your-master] Copying the style of an example and honouring a corporate master are two different jobs. The first is about tone and rhythm. The second is about placeholders, layouts, theme colours, fonts, protected pages, and whatever your brand team checks before anything goes out. The official add-in writes into text placeholders and native shapes, which is the right way to build it. But OpenAI's [help documentation](https://help.openai.com/en/articles/20001242-chatgpt-for-powerpoint) says template adherence and advanced chart, shape, formatting, and slide-management edits can still be limited, and the [beta launch](https://www.thurrott.com/a-i/336411/openai-launches-chatgpt-for-powerpoint-add-in-in-beta) FAQ names template and font handling specifically. OpenAI also asks users to review every change the add-in makes. The layer that turns a good answer into a master-compliant slide is still being built, and it is the hard part. The [native slides explainer](/en/blog/editable-ai-powerpoint-native-slides) explains why master inheritance, chart objects, and text autofit are where generated decks usually break. ### 4. Nothing tells the model what it may use [#4-nothing-tells-the-model-what-it-may-use] A professional deck comes with rules. Certain pages cannot change, certain claims need sign-off, certain data cannot leave a defined environment, and certain figures have to trace back to an approved source. Astra knows none of this unless you build it around the model. There is no locked disclaimer page, no approved claims register, no reference that may only be shown to some clients. ChatGPT's workspace settings control what the model can access. They do not control what ends up on slide 14. The [AI agents in PowerPoint guide](/en/blog/ai-agents-powerpoint-enterprise-guide) puts it plainly: you control an agent through the tools and permissions it runs under, not through a better prompt. Astra is a more capable agent. Designing the permission model is still your job. ## A 30 minute test on your own deck [#a-30-minute-test-on-your-own-deck] Do not judge this on OpenAI's demo. Take a proposal or board pack your team finished last month, strip out the final version, and give Astra the same inputs the team had. * **Provide the actual master, not a clean template.** Request a ten slide deck. Open the result in outline view and confirm whether content sits in placeholders or in free text boxes. * **Include one internal reference.** A project credential or person profile that exists only on the intranet. Record whether it appears, is invented, or is silently omitted. * **Request a figure that changed last quarter.** Confirm whether the deck uses the current number, the previous one, or a plausible substitute. * **Require a source on every chart.** Open each source and confirm that it exists and supports the specific figure. * **Change a theme colour in the master afterwards.** Confirm whether the generated deck follows or the colour is hard coded. * **Ask for an edit to one slide only.** Compare the file before and after. Confirm that nothing else changed. * **Add a locked disclaimer page and request a restructure.** Confirm the disclaimer survived unchanged. * **Pass the result to the brand reviewer without context.** Record the number of corrections. * **Measure the review time, not the generation time.** Generation takes a minute. The relevant measure is how long a senior reviewer needs before signing off. The last item is the one that counts. With any frontier model, generating the deck is now close to free. What you pay for is checking it, and that cost depends on the workflow, not the model. ## When Astra on its own is enough [#when-astra-on-its-own-is-enough] | Situation | Astra on its own | Astra with a governed layer | | -------------------------------------------------------- | ----------------------------- | --------------------------------------------------- | | Team update, training deck, internal status review | Sufficient | Not required | | Rewriting, shortening, or restructuring an existing deck | Sufficient | Not required | | Internal working draft on a public topic | Sufficient | Not required | | Storyline and structure before analysis | Sufficient | Not required | | Client proposal with credentials and references | Draft only | Required for approved content and master compliance | | Board or investment committee pack | Draft only | Required for source trail and locked pages | | Regulated or confidential material | Depends on workspace controls | Required for data boundaries and audit trail | | Recurring decks from live data | Manual each cycle | Required for repeatable, reviewable runs | The top four rows are most of what a team produces in a normal week, and ChatGPT covers them. [Human review](/en/blog/human-in-the-loop-ai-presentations) applies in every row. The difference lower down is how much of that review a system has to make cheap, and that is the job of a governed layer. ## Where offgen fits [#where-offgen-fits] We do not compete with Astra on intelligence, and I am glad the general tools have become this good. Every model release helps our users, and a team that already works comfortably with ChatGPT will find governed generation familiar. What offgen adds is the depth this article is about. Generation runs against your approved [Company Brain](/en/blog/powerpoint-company-brain-guide), meaning reusable slides, people records, project references, and the wording your firm has signed off. Output is [native PowerPoint](/en/product/editable-templates) mapped to your master's placeholders, so it survives a theme change and a brand review. [Lockable elements](/en/product/lockable-elements) protect pages that must not move, and [brand governance](/en/product/brand-governance) checks every generated slide against your rules. Processing stays in the EU. If your teams run on ChatGPT, the two fit together. offgen exposes its workflows through [MCP](/docs/mcp), so an external AI client can start a governed generation run and get the finished file back. Our [ChatGPT for PowerPoint comparison](/en/alternatives/chatgpt-powerpoint) lists the differences, including where the general assistant is the better pick. Astra is the best model to date for thinking about a deck, and for everyday PowerPoint work you should use it without hesitation. For the decks that carry your name to a client or a board, ask whether the workflow around it knows what your firm knows, respects your master, and leaves a trail a reviewer can follow. Test that on your own deck before you decide. ## Frequently asked questions ### Can GPT-6 Astra create PowerPoint slides directly? Yes, in two ways. Inside ChatGPT it can produce presentation files as part of its professional work capabilities. Inside PowerPoint, the official ChatGPT add-in, which entered global beta in May 2026, drafts and edits native slides from a side panel. Both paths deliver editable output, but OpenAI documents that template and font handling and advanced chart and formatting edits can still be limited. ### Is Astra better at slides than earlier models? OpenAI says it is its best model yet for adhering to existing templates and producing well laid out slides with a structured narrative. Independent testing by Artificial Analysis found large gains in analytical quality and a lower hallucination rate, but a lower presentation quality score than GPT-5.6 Sol. So the answer depends on whether you mean the thinking on the slide or the look of the slide. ### Does Astra do enough research for a consulting deck? Not on its own. It browses the public web well and cites what it finds, but a professional deck rests on approved internal knowledge, licensed data, project references, and current figures that never appear on the open web. Astra cannot reach those unless you connect them, and it does not carry a source trail onto the slide that a reviewer can verify. ### What are the biggest gaps for professional teams? Four recur in our testing and in OpenAI's own documentation: research is limited to what the model can reach, master and layout fidelity is not guaranteed in the PowerPoint add-in, no citation or approval trail is attached to the generated slide, and there is no brand governance or locked element control. Each one is manageable with process, but none of them is fixed by a smarter model. ### Should a consulting or finance team stop using ChatGPT for slides? No, quite the opposite. ChatGPT and the PowerPoint add-in cover everyday needs well: team updates, training decks, internal reviews, rewrites, restructures, and first passes on content. Keep using them for that. For client proposals, board packs, and regulated material, route the deck through a governed layer for approved content, master compliance, citations, and review before release. ### How often should this assessment be rechecked? Every 90 days at most. OpenAI ships model and add-in updates faster than any review cycle, and the beta status of the PowerPoint add-in means capabilities documented today may change without notice. ## Sources 1. [GPT-6 Astra: A new generation of intelligence](https://openai.com/index/gpt-6-astra/) — OpenAI, 2026-09-03; accessed 2026-09-11. 2. [GPT-6 Astra: The next generation in intelligence for work](https://openai.com/index/gpt-6-astra-next-generation-work/) — OpenAI, 2026-09-03; accessed 2026-09-11. 3. [GPT-6 Astra System Card](https://deploymentsafety.openai.com/gpt-6-astra) — OpenAI Deployment Safety Hub, 2026-09-03; accessed 2026-09-11. 4. [Benchmarking GPT-6 Astra](https://artificialanalysis.ai/articles/benchmarking-gpt-6-astra) — Artificial Analysis, 2026-09; accessed 2026-09-11. 5. [ChatGPT for PowerPoint](https://help.openai.com/en/articles/20001242-chatgpt-for-powerpoint) — OpenAI Help Center; accessed 2026-09-11. 6. [OpenAI Launches ChatGPT for PowerPoint Add-in in Beta](https://www.thurrott.com/a-i/336411/openai-launches-chatgpt-for-powerpoint-add-in-in-beta) — Thurrott, 2026-05-22; accessed 2026-09-11. 7. [ChatGPT for PowerPoint alternative](https://www.offgen.ai/en/alternatives/chatgpt-powerpoint) — offgen; accessed 2026-09-11. 8. [Editable templates](https://www.offgen.ai/en/product/editable-templates) — offgen; accessed 2026-09-11. 9. [MCP documentation](https://www.offgen.ai/docs/mcp) — offgen; accessed 2026-09-11. ## Related articles - [Editable AI PowerPoint: Why Native Slides Matter for Professional Teams](https://www.offgen.ai/en/blog/editable-ai-powerpoint-native-slides) - [AI Agents in PowerPoint: Secure Workflows, Controls, and Use Cases](https://www.offgen.ai/en/blog/ai-agents-powerpoint-enterprise-guide) - [PowerPoint Citations: Source Footers, Links, and AI Verification](https://www.offgen.ai/en/blog/powerpoint-citations-source-footers) - [What Is a PowerPoint Company Brain? An Enterprise Guide](https://www.offgen.ai/en/blog/powerpoint-company-brain-guide)