Which AI Model Actually Makes the Best Slides

I gave three AI models the exact same slide brief and watched them fail in three different ways. Then I watched them get better in three different ways too. That’s the more interesting part.

The brief was simple on purpose: a 10-slide investor update for a fictional B2B SaaS company. Revenue growth, churn, a roadmap slide, a team slide, an ask. No design instructions, no formatting hints, no “keep it minimal” nudging. I wanted to see what each model reaches for by default when nobody’s watching.

I ran the same prompt through Claude Sonnet, GPT-4o, and Gemini, and fed every output into Slaide as a .slaide file so I could compare them stripped of any styling bias. Same theme, same font, same everything except the content each model chose to write. That turned out to matter a lot.

The Text Density Problem

Here’s the thing that separates a slide from a document: word count per slide. Every model I tested has read a thousand blog posts and PDFs in training, and it shows. The default instinct, unless you fight it, is to write paragraphs.

GPT-4o was the worst offender here. Its churn slide had four full sentences and a sub-bullet with three more clauses. Read out loud it took 40 seconds. That’s not a slide, that’s a memo with a title.

Gemini did better on density but overcorrected into fragments that lost meaning. “Churn down. Retention strong. Q3 outlook positive.” Three lines, six words total, and I genuinely couldn’t tell you what changed or why. Density isn’t the only goal - the words still have to carry information.

Claude landed in the middle and, honestly, in the right place. Its churn slide: “Churn dropped from 4.2% to 2.8% monthly, driven by the onboarding rewrite in May.” One line. Numbers doing the work. That’s the sentence a human would actually say out loud while pointing at the slide, not read verbatim from it.

Structure: Who Understands a Deck Has a Shape

A good deck isn’t 10 independent slides. It has a shape - open with the hook, build through evidence, land on the ask. I wanted to see if any model tracked that across the whole brief or treated each slide as an isolated task.

GPT-4o treated them as isolated. Slide 3 repeated a stat from slide 2 almost verbatim, because nothing in its process referenced what came before. It’s a known failure mode with long single-shot generations - the further you get from the start of the response, the less the model seems to remember what it already said.

Gemini was better here, and this was the surprise. It built a real narrative arc without me asking for one. The roadmap slide explicitly called back to the churn number from three slides earlier (“this roadmap addresses the drivers behind that 2.8% number directly”). I didn’t ask for cross-referencing. It did it anyway, and it’s the kind of thing a good human deck author does instinctively.

Claude’s structure was solid but more conventional - each slide clean and self-contained, correct ordering, no repetition, but no unprompted callbacks either. Competent, not memorable.

Visual Hierarchy Without Being Told

None of the three got a single word about visual weight, headline sizing, or what should be big versus small. I wanted to see what each model assumed a slide title even is.

Claude and Gemini both defaulted to short, declarative titles: “Churn Is Down,” not “An Overview of Our Customer Retention Metrics for the Quarter.” GPT-4o wrote titles that were themselves full sentences half the time, which meant when Slaide rendered them at title size, they wrapped to two lines and ate a third of the slide before the body text even started.

That’s not a cosmetic complaint. A wrapped title pushes everything else down, and now your bullet points are competing for a shrinking bottom half of the canvas. The model that writes a six-word title is quietly doing layout work it was never asked to do.

Where GPT-4o Won

I don’t want to make this sound like a clean sweep, because it wasn’t. GPT-4o’s team slide was the best of the three by a clear margin. It suggested a 2x2 grid with role, one metric, one line of context per person - a layout none of the other models proposed unprompted. When it wasn’t fighting its own verbosity, its instinct for “what layout serves this specific content” was sharp.

That’s the pattern across a few dozen decks I’ve now run through this comparison, not just the one brief: GPT-4o’s structural instincts are good, its restraint is the problem. If you tell it explicitly “one sentence, max 12 words” it snaps into line immediately. Left to its own judgment, it drifts long.

What This Means If You’re Picking a Model

If I had to rank them for slide generation specifically, today, with no prompt engineering: Claude first for the combination of density discipline and title hygiene, Gemini second for genuine structural awareness across a deck, GPT-4o third purely on the verbosity problem - its per-slide creative instincts are honestly excellent when constrained.

None of this is fixed forever. These are snapshots of three models on one day in August 2026, and model behavior shifts with every release. The gap between GPT-4o unconstrained and GPT-4o with a strict word cap was bigger than the gap between any two models left unconstrained. So the real lesson isn’t “use model X” - it’s “write your prompt like you’re briefing a junior who’s never seen a slide deck before,” because apparently that’s closer to true than I expected.

How I Actually Ran This

I built Slaide because I got tired of AI-generated decks that looked like Word documents wearing a slideshow costume, and this comparison is exactly the kind of thing it’s for. Every model’s output went through the same Slaide MCP server, writing .slaide files directly - plain text markup, git-diffable, no PowerPoint XML in sight. The core is Apache-2.0, so if you want to run this same three-model comparison yourself, the tooling is right there.

One honest limitation: none of this tests whether the models can iterate. I gave one brief, took one output, compared cold. In practice you’d go back and forth - “tighten slide 4,” “the team slide needs a different layout” - and that conversational loop probably narrows the gap between all three more than the single-shot numbers above suggest.

Still. If you’re piping raw model output straight into a deck with no human editing pass, the model you pick matters more than I expected going in.