How to Use GPT-6 Astra to Create Research-Oriented Slide Decks

Combine those into a single throughput figure and the gap is larger than either number suggests on its own:

Successful tasks per hour = (60 ÷ minutes per task) × success rate. GPT-5.6 Sol: (60÷75) × 0.657 = 0.53. GPT-6 Astra: (60÷40) × 0.726 = 1.09. Derived from OpenAI’s published OSWorld 2.0 figures.

Astra clears 2.1× the successful agentic work per hour that its predecessor did. For a model whose job is to read your paper, pull your references, and assemble twenty slides against your department template, throughput per hour is not an abstract benchmark. It is the difference between a tool you use before a lab meeting and a tool you abandon halfway through one.

Why Research Decks Are the Hard Case

Most AI slide tools are tuned for a pitch deck: eight slides, big claims, no sources. A research deck inverts every one of those assumptions.

The source material is long and structured. A paper is thirty pages of methods, tables, and figures, not a topic string. A model that only accepts a prompt box cannot see your actual results.

Claims need attributions. A slide that says “outcomes improved at 12 months” without a reference is unusable in a lab meeting, a grand rounds, or a thesis defense. Citations are not decoration in academia; they are the load-bearing element.

Structure is argued, not templated. A conference talk holds one argument across sections. Weak models flatten distinct findings into one bullet, or generate twelve slides that restate the same point in different words. TechForum’s walkthrough of how to make journal club slides with AI makes the same point from the other direction: the format only works when the deck carries a single critical argument about the paper, rather than a section-by-section recap of it.

This is exactly where a stronger reasoning model helps, and exactly where model choice alone is not enough. You need a tool that can feed the model your PDF and retrieve real references, and then render the result into slides that look like your field expects.

The Workflow: PDF In, Cited Deck Out

Here is the sequence that works, using ChatSlide, which added GPT-6 Astra to its model picker on September 4, 2026, one day after OpenAI’s announcement, making it one of the first presentation applications to ship the model in production.

1. Upload the paper, not a summary of it. Drop in the PDF directly. ChatSlide extracts the text and tables, including OCR on scanned documents, so the model reasons over your actual methods and results rather than over its general knowledge of your field. This single step is what separates a deck about your topic from a deck about your paper.
2. Pull real citations from scholarly databases. ChatSlide supports direct PubMed search and research imports from sources including OpenAlex, CrossRef, and ClinicalTrials.gov. arXiv is not currently a built-in search source, although arXiv content can be added through the URL pipeline. This helps keep the references on your slides tied to real source records rather than model-generated approximations of paper titles. For anyone who has watched a language model confidently invent a plausible-sounding reference, this is the feature that makes AI-assisted academic work more defensible.
3. Select Astra before you generate. The model picker sits at the top of the Summary step. Free accounts run GPT-5.6 Luna, the Plus plan adds Terra, and Astra sits on the Ultimate tier, described in-product as the strongest reasoning option for demanding content. Model choice is per deck, not an account-wide setting, so you can draft on a fast model and regenerate the same deck on Astra when a section needs more depth. Your brief, uploads, and outline carry over.

The picker in production, September 2026: Luna on the free tier, Terra on Plus, and GPT-6 Astra on Ultimate.

4. Write a brief that states the argument. Say what the talk must establish, who is in the room, and what they should conclude. “Summarize my paper” produces a table of contents. “Convince a skeptical methods reviewer that our replication holds across all three runs” produces a talk.
5. Fix the outline before slides exist. ChatSlide generates the structure first. Repairing the argument at outline stage costs seconds. Repairing it after twenty slides are designed and captioned costs an afternoon.
6. Generate, refine, export. Charts build from your uploaded spreadsheet with real axes and real proportions. Nineteen AI editing tools handle the after-work: compress an overloaded slide, turn a finding into a recommendation, generate a diagram. Per-slide speaker notes size to your actual speaking slot, which for a fifteen-minute conference talk is the difference between finishing and being cut off. Export to PowerPoint, PDF, or Keynote, or render the deck as a narrated video for collaborators who cannot attend.

Which Tools Can Actually Run GPT-6 Astra Today

This is where the field separates, and it separates sharply.

Astra rolled out in phases: a limited set of organizations first, then ChatGPT Plus, Pro, Business and Enterprise, then the OpenAI API and AWS. Any application offering it needed API access lined up in advance.

Competitor rows reflect publicly documented capabilities as of publication. None of these vendors had publicly announced GPT-6 Astra availability at the time of writing.

Two gaps in that table matter more than the Astra column itself.

Web search is not the same as scholarly retrieval. Gamma’s 3.0 Agent can research the web and attach citations, which is genuinely useful for a market-sizing deck. It is a different instrument from an indexed literature search. A web query can surface a press release summarizing a trial, a secondary blog post, or a preprint that was later retracted, and cite any of them with equal confidence. Querying PubMed, arXiv, OpenAlex, or Semantic Scholar returns the indexed record itself, with the identifiers a reviewer uses to verify it. For a marketing deck the distinction is academic. For a thesis defense the distinction is the whole thing.

Model choice is rarer still. Most tools pick a model for you and never tell you which one. When a better model ships, you wait for the vendor to swap it silently, and you cannot choose depth over speed for the one deck that needs it. A visible picker is what let ChatSlide users move to Astra on day one instead of waiting for a migration.

What a Better Model Will Not Fix

Two limits are worth stating plainly, because launch coverage tends to skip them.

A stronger model does not replace the presentation layer. Astra can improve template adherence, structure, and slide layout, but typography, chart rendering, themes, and export fidelity still depend heavily on the presentation tool and template.

A frontier model does not fix a vague brief. “Make me a deck about my paper” produces a more elaborate wrong deck on Astra than on a weaker model. The brief remains the highest-leverage input, and no model release will change that.

The Practical Takeaway

GPT-6 Astra is a flagship model for which OpenAI explicitly highlighted presentation generation and template adherence as part of professional work, and its agentic throughput gain, 2.1× the successful tasks per hour of its predecessor, is derivable from OpenAI’s published OSWorld 2.0 figures.

For research decks specifically, the model is necessary but not sufficient. What makes it usable is the surrounding machinery: PDF ingestion so the model sees your actual paper, scholarly database retrieval so your citations are real, and a visible model picker so you can spend the stronger model on the deck that warrants it.

Researchers evaluating tools for conference talks, thesis defenses, and lab meetings can find ChatSlide’s academic workflow, including its citation and PDF pipeline, at its AI slide maker for academia

Quanlai Li is the founder of ChatSlide, an AI presentation platform used across academia and industry.