The Best LLM for Creating Presentations

It’s 8:45 and you realize you spaced that presentation deck you were supposed to make for your nine o’clock. What do you do? You could Slack your designer a 9-1-1, but they’re already at capacity — and what could they do in 15 minutes, anyway?

Looks like you’re on your own.

Or are you?

AI to the Rescue

Major large language models (LLMs) can now create full, multi-slide presentations from a simple prompt. There’s little you need to do other than supply a topic. But are the results usable?

We compared three LLMs (Anthropic’s Claude Sonnet 5, OpenAI’s ChatGPT 5.5, and Google’s Gemini 3.5 Thinking) to see which is the most qualified to assist you when you need a slide deck ASAP.

Our first test was designed to gauge the basic design chops of the models. We provided detailed instructions, including verbatim copy, for three (very rudimentary) slides. The goal was to see how each LLM handled general layout, copy placement, simple data visualization, and custom illustrations.

For our next test, we provided a simplified prompt that provided no verbatim information; just a topic, a limit on the number of slides, and a request for relevant graphs or illustrations. Then we let the LLMs run wild. This tested the models’ ability to generate and summarize information and create a compelling presentation without any handholding from us.

For both tests, we uploaded a PDF of the Devise brand guidelines to give the models some basic creative direction. Each LLM was also prompted to deliver the presentation as a PowerPoint file.

Test 1

Prompt: Create a 3-slide presentation that includes copy, data, and illustrations as directed below. All slides should have a uniform look and feel and use the given color and typography information.

Primary color: #F26800
Secondary color: #191C1A

Headline typeface: Inter 700

Body typeface: Inter 400

Slide 1:

Title: A very exciting presentation

Body: Let’s see how this LLM handles a basic copy slide. Here it is. How’s it look?

Slide 2:

Title: Very important data

Body: Now let’s see how well it handles creating a graph from some data.

Illustration: Build a bar graph with “$” on the Y axis and “Years” on the X axis that visualizes the following data:

Year 1 / $20

Year 2 / $15

Year 3 / $40

Slide 3:

Title: Illustration of an abstract concept

Body: Nothing makes a quality presentation like an on-point illustration. Let’s see how this LLM visualizes the concept of creativity.

Illustration: Create an illustration that demonstrates the concept of creativity.

Google Gemini

Gemini created a new presentation in the form of an HTML document rather than the requested PowerPoint file, but it included an animated element — the only model to do so.

While all three models pulled our company name from the brand guidelines, Gemini also found our tagline (“Insightful Imagination”) and applied it to the slides. That was a nice touch. Again, we did not provide specific instructions on how the models should use the brand guidelines, so Gemini’s decision-making stands out here.

Where Gemini fell short was the graph, in which the height of the bars is the same on all values and not accurate for any. Interestingly, in a previous attempt, it did not make this mistake but did flip the X/Y axes despite our clear instructions.

Gemini, in our opinion, also created the most visually appealing layout.

OpenAI ChatGPT

ChatGPT was the only model to go rogue and add its own copy. It’s in that white box on the right side of Slide 1. It’s incredibly pointless, but then again, we didn’t exactly give it anything real to work with. Perhaps this sort of “creativity” would help in a real presentation?

ChatGPT nailed the graph, but it also included the data values we provided as separate copy on the slide. We didn’t want that, but we also didn’t specifically say we didn’t want it in the prompt. More detail in the prompt would likely solve this.

The image ChatGPT generated (or found) to illustrate “creativity” is nothing exceptional, but we like it more than Gemini’s atom diagram.

Anthropic Claude

Claude is a bit of a mixed bag. The overall layout is dull, but clean and easy on the eyes. We’d put it ahead of ChatGPT but behind Gemini. It also added a nonsense illustration to the first slide, all of its own accord.

While none of the models impressed us with their interpretation of “creativity,” Claude’s “lightbulb” and unprompted caption (‘Ideas in motion’) are groan-worthy. However, Claude gets points for making every element of its graphics editable. All those little shapes surrounding the lightbulb? You can drag, resize, and rotate those to your heart’s content. That’s pretty neat!

Test 2

Prompt: I need to create a presentation on the evolution of creative advertising from 1920 to today. Please perform all applicable research and condense it into a maximum of 5 slides. Generate and include graphs from any relevant data. Include other illustrations as needed. Use the provided brand guide PDF to inform the colors, fonts, and other styles of the presentation deck. Save the final presentation in a PowerPoint (.pptx) file.

Google Gemini

Once again, Gemini put together the most visually appealing deck. Images, concise copy, and a large (and actually accurate) graph all came together to make this the most “professional” result.

It also did something unique with Slide 5 — it made it about us, a sort of call-to-action for Devise. It wasn’t in the ask, but we like that Gemini at least tried to accomplish something with this presentation.

Gemini was the only model that didn’t adhere to our five-slide limit. However, because the sixth slide was purely a list of references, we’ll let it slide. No pun intended.

OpenAI ChatGPT

ChatGPT delivered a modern and striking layout, but the illustrations are a bit wonky and don’t quite land for us. It’s also the only model that used very small text sizes for some secondary copy, which would likely be illegible on some screen sizes. Fortunately, at least all the primary copy was appropriately sized.

Anthropic Claude

Bullet points, bullet points, bullet points. Eh, sorry, Claude, but you put together the most boring layout of the three. On the plus side, it probably crams the most overall information into the slides, but sadly they’re almost painful to actually look at.

We do like the graph on page three that shows the takeover of digital. And it’s nice how it used all our brand colors there.

Conclusion

I think we first need to state the obvious: none of the results here really hold a candle to a hand-produced presentation. But when time is short and options are limited, AI can churn out something that isn’t half bad.

Overall, we were impressed by each model’s ability to make use of our brand guidelines, at least with regard to colors, and put together passable presentations with little guidance from us. In our second test, all models did a good job of presenting data. (Of course, we don’t know if it’s actually accurate.) Gemini gave us our preferred results in both tests, but it didn’t always nail everything — the bad graph in the first test was pretty bad.

Many of the issues we encountered across models could likely have been fixed with a better prompt. This, we suspect, is more important than the model itself. We’re sure any of the LLMs tested could produce better results, especially if you take the time to test and experiment.

That said, when it comes to throwing something together last minute, we’d trust Gemini over the others thanks to its clean styling and strong layouts.