1
claude-opus-4.8
Anthropic
4.8/ 5 overall
Best at storyline
Which AI writes the best decks?
Text models only, ranked on how well they follow the unPaper prompt and how good the resulting deck becomes.
Edition 2026-07-15 · 6 makers · 16 runs
The podium
2
gpt-5.2
OpenAI
4.8/ 5 overall
Best at titles tell the story
3
grok-4.5
xAI
4.6/ 5 overall
Best at conformance
Ranked models
All tested models
Scores are means across every filed run. Bad runs count.
| Rank | Model | Self-ID | |||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
1 | claude-opus-4.8· api Anthropic · reasoning off · 2 runs | 4.8 | 5.0 | 5.0 | 5.0 | 4.5 | 4.0 | 5.0 | 5.0 | 4.5 | 100% | 8.0 | 0/2 |
2 | gpt-5.2· api OpenAI · reasoning on · 2 runs | 4.8 | 4.5 | 5.0 | 4.5 | 4.5 | 5.0 | 5.0 | 5.0 | 4.5 | 100% | 9.0 | 2/2✓ |
3 | grok-4.5· api xAI · reasoning on · 2 runs | 4.6 | 4.5 | 4.5 | 4.5 | 4.5 | 5.0 | 5.0 | 5.0 | 4.0 | 100% | 8.0 | 0/2 |
4 | deepseek-v4-pro· api DeepSeek · reasoning not recorded · 2 runs | 4.1 | 4.0 | 4.5 | 3.5 | 4.0 | 3.0 | 5.0 | 4.5 | 4.0 | 100% | 7.0 | 0/2 |
5 | gemini-3.1-pro-preview· api Google · reasoning on · 2 runs | 4.1 | 3.5 | 5.0 | 3.5 | 4.0 | 4.0 | 5.0 | 4.0 | 4.0 | 100% | 7.0 | 0/2 |
6 | gpt-5-mini· api OpenAI · reasoning on · 2 runs | 4.1 | 4.0 | 4.5 | 4.0 | 3.0 | 3.0 | 5.0 | 5.0 | 4.0 | 100% | 9.0 | 0/2 |
7 | mistral-large-2512· api Mistral AI · reasoning off · 2 runs | 3.9 | 3.0 | 4.0 | 3.5 | 4.0 | 4.0 | 5.0 | 5.0 | 3.0 | 100% | 10.0 | 0/2 |
8 | deepseek-v4-flash· api DeepSeek · reasoning off · 2 runs | 3.6 | 3.5 | 3.5 | 4.0 | 3.0 | 2.0 | 5.0 | 4.5 | 3.5 | 100% | 7.0 | 0/2 |
Dimension leaders
Story · claude-opus-4.8Titles · claude-opus-4.8Pattern · claude-opus-4.8Visual · claude-opus-4.8Confm · gpt-5.2Convert · claude-opus-4.8Native · claude-opus-4.8Source · claude-opus-4.8
The models under test
DeepSeek
DeepSeek
4.1 / 5 · deepseek-v4-pro
Mistral
Mistral AI
3.9 / 5 · mistral-large-2512
Gemini
4.1 / 5 · gemini-3.1-pro-preview
Grok
xAI
4.6 / 5 · grok-4.5
ChatGPT
OpenAI
4.8 / 5 · gpt-5.2
Claude
Anthropic
4.8 / 5 · claude-opus-4.8
How the score is built
Each text model gets the real unPaper prompt and the same brief. The saved HTML and converted PowerPoint are scored 1 to 5 on eight fixed dimensions. Overall is the mean. Conformance is auto-scored; the rest are hand-scored against published anchors. The sample is deliberately small, so runs are shown and no confidence intervals are invented.
Self-ID counts the runs where the model, asked to stamp a machine-readable note naming its generator, named itself plausibly — hover the cell for what each model actually called itself. Runs driven over an API carry no system prompt telling the model who it is, so this column measures self-knowledge, not deception.
Variety counts how many distinct exhibit families a deck uses — charts with editable data, real tables, stat rows, matrices, flows, timelines, scorecards, quotes, dividers, vector graphics, layout grids and the signature consulting exhibits — detected mechanically, out of 14. It measures compositional range, not correctness; the Pattern-fit score judges whether the choices were right.
Every row ships its receipts: expand a model to open the exact deck it produced — untouched — and the converted PowerPoint. The full prompts the models received: 2026-07-15-exec-macro-briefing · 2026-07-15-market-entry.
- StorylineStory
- Answer first, grouped support, a closing ask — not a content dump.
- Titles tell the storyTitles
- Read only the titles: full-sentence findings, not labels.
- Pattern fitPattern
- The right element per idea — table for a comparison, chart for a trend.
- Visual disciplineVisual
- One accent, visual anchors, a signature moment, nothing crowded.
- ConformanceConfm
- Verbatim style/@page/script, look untouched, notes on every slide, complete file. Auto-checked.
- ConversionConvert
- Share of the deck that converts to native PowerPoint objects, not images.
- Native editabilityNative
- Everything movable, charts data-editable, theme on-brand in PowerPoint.
- Source disciplineSource
- Numbers correct and sourced, no hallucinated precision.
Generated 2026-07-15 · 16 runs · bad runs count too