# Presentations test — 2026-09

**Brief, sent once to each tool with the memo pasted underneath** (`fixtures/brief.md`, `fixtures/q2-memo.md`):

> Make me an 8-slide deck for my two-person team about our shop's Q2 results, using only the facts in the memo below - do not add any. Slides: 1 title, 2 revenue by month, 3 revenue by category, 4 top products, 5 orders by region, 6 the two data problems we found, 7 three actions for Q3, 8 closing. Short speaker notes on every slide. Give me the deck as a .pptx file, or a link I can open in PowerPoint or Google Slides.

Run by Claude Code in the operator's Chrome session on 2026-09-17, 10:16–10:29 (+07), with Kimi's answer at 16:25 after hourly retries, and the two deck-builder Specialists, Gamma and Canva, on 2026-09-18 at 20:26–20:48, free tiers, fresh chats, Claude in an incognito chat. The memo is seven bullets about a fictional Lisbon shop's quarter - 25 checkable facts, built from the Excel test's cleaned data - so every number on every slide has a right answer. Each returned deck was opened with python-pptx (`fixtures/check_deck.py`): slide count, chart data, notes per slide, facts found; and checked structurally (`fixtures/check_layout.py`). The decks are in `fixtures/returned/`.

**Measured figure: a deck delivered or not, then facts of 25 and notes of 8.** A deck is a .pptx that opens or a link that opens in PowerPoint or Google Slides; an outline in chat is not a deck.

| | ChatGPT | Claude | Gemini | Perplexity | Copilot | DeepSeek | Grok | Qwen | Kimi | Gamma | Canva |
|---|---|---|---|---|---|---|---|---|---|---|---|
| Deck delivered | **refused** - data-analysis limit until tomorrow | **.pptx**, ~10 min | HTML canvas, ~32 min; its Export → PPTX works | **.pptx**, 19 s | **.pptx**, ~2 min | no - text outline | **.pptx**, 4m 32s | no - text outline | **.pptx**, ~4 min - on the sixth attempt, after six hours of "too many people are talking to Kimi" | **.pptx** via Export, ~4 min | **no** - free plan cannot take the brief |
| Slides | — | 8 | **9** (+ "Image Sources") | 8 | 8 | 8 in text | 8 | 8 in text | 8 | 8 | — |
| Facts (of 25) | — | 25 | 25 | 25 | 25 | 25 | 25 | 25 | **24** - "20 customers" left out | **16** as text; 5 more only inside a chart picture, 2 rounded there, 2 absent | — |
| Invented | — | 0 (two derived lines, both true) | **2**: a web photo of a real shop, and "all 5 regions reached 120 orders" (they sum to 118) | 0 | 0 | 0 | 0 | 0 | 0 (one embellishment, "10-minute review") | **3**: a monthly average, category percentages, "Ceramic Bowl - top Home product" | — |
| Notes (of 8) | — | 8 | **0** in the notes pane - printed on the slide face instead | 8 | 8, three to five words each | 8, as text | 8 | 8, as text | 8, the fullest of the nine | **0** in the notes pane - callout boxes on the slide face | — |
| Charts | — | 3 bar charts with the memo's values | none (pictures on every slide) | none | 4 charts | — | none (number tiles) | — | 4 charts | none in the file - 4 charts exported as pictures | — |
| Look | — | designed: title slide, cards, footer with page numbers | designed, with a stock photo it found itself | bare: two text boxes per slide, 4:3 | default theme, one chart per slide | — | designed from shapes: header strip, page numbers, tiles | — | designed: green title, section labels, side panels | designed: AI illustrations, tiles, cards, a Made with Gamma badge | — |
| Output form | one notice | file card + in-chat preview | canvas with export menu | artifact card | download link | chat | file card | chat | file card + preview | editor + Export menu | an error, then a Pro button |
| Quality / Effort / Cost / Free tier (0–3) | 0/0/3/0 | 3/3/3/3 | 1/2/3/3 | 2/3/3/3 | 2/3/3/3 | 1/1/3/3 | 3/3/3/3 | 1/1/3/3 | 3/3/3/1 | 1/3/3/2 | 0/0/3/0 |
| What stood out | Same limit as the spreadsheet - no code, no deck. | The only deck you would present as it is. | Added a photo of somebody else's shop and a number that doesn't add up. | The fastest deck, and the plainest. | Real charts in two minutes; notes you'd have to write yourself. | Honest about the file, then everything else. | A designed deck with every number in its place. | The one that told you about Slides from Outline. | The best notes and four charts, six hours late. | The prettiest deck and the least PowerPoint - every chart is a picture, the notes are on the slides, and it added numbers. | No Canva AI on the free plan; a 280-character prompt box; Convert is Pro. |

## Notes per tool

- **ChatGPT.** Free plan. The reply was empty apart from the notice that the data-analysis limit is used up until 09:13 the next day. The same allowance a deck needs; it had already gone on the Excel test. Retried the next morning at 09:25, after the reset hour: "It seems like I can't do more advanced data analysis right now" - the Excel retry three minutes earlier had used the day's allowance again.
- **Claude.** Sonnet 5, Medium, incognito. About ten minutes of visible work ("Read 9 files, ran 6 commands"), then "Quinta da ribeira q2 2026.pptx" with a preview. Burgundy title slide; bar charts on the month, category and region slides with the memo's numbers as the chart data; numbered cards for products, problems and actions; a footer with the shop name and page number on every slide; notes on all eight. Two lines that are not in the memo but follow from it: category shares "41 / 33 / 26%" (the notes call them approximate) and "Lisboa's orders were worth more on average". Smallest text 10 pt in footers and chart labels.
- **Gemini.** Flash. Chose an HTML "presentation" canvas rather than a file; the card said "Генерация…" for half an hour and then stamped 10:52. The canvas renders well - serif title, four stat tiles, cards - and its Export menu produced a real PPTX. But: nine slides, the last one "Image Sources" for a photograph it fetched from oladaniela.com ("Best shops Évora") and put on the deck; the speaker notes are typed onto the face of each slide and the exported file has none in the notes pane; and slide 5 says "Combined order total across all 5 regions reached 120 orders" when the memo's five counts add up to 118. Filler on the action slide ("to leverage top product momentum").
- **Perplexity.** Nineteen seconds to an artifact and a download. Every fact, notes on every slide, eight slides in the order asked - as two plain text boxes per slide on a white 4:3 page with no set font sizes and no charts. The artifact menu also saves straight to Drive as PowerPoint or Google Slides.
- **Copilot.** Two minutes, "Download the PPTX deck". Four real charts (months, categories, products, regions) carrying the right values, the other slides as short bullets, default theme, 22–24 pt. The notes exist on all eight slides and say almost nothing: "Quarter results overview.", "Revenue by category.", "Issues identified and fixed."
- **DeepSeek.** "I can't create or host a .pptx file or a shareable link directly in chat." Then the eight slides as text with notes under each, all 25 facts, and instructions to paste them into Google Slides. Adds one line of cheer to the closing notes.
- **Grok.** "Работало на протяжении 4m 32s", then a file card and an outline. The deck is built from shapes: a header strip with the shop name, "INTERNAL · TWO-PERSON TEAM" and "0n / 08" on every slide, big-number tiles for months, categories, products and regions, numbered actions. No charts. Notes on all eight. The in-chat preview opened a file browser (slides/, unpacked/) and a spinner instead of the slides.
- **Qwen.** "As an AI, I cannot directly generate a downloadable .pptx file." The outline it gives is formatted for PowerPoint's Home → New Slide → Slides from Outline import, which would build the eight slides from a .txt; the notes would still be pasted by hand. All facts, nothing invented.
- **Kimi.** The capacity modal on send at 10:29, and again at 12:24, 13:33, 14:24 and 15:24. At 16:25 it accepted the brief, ran python-pptx three times, noticed "Slide 8's notes didn't save", fixed it, and delivered in about four minutes: eight 16:9 slides with four real charts, a section label and number on each, side panels for the headline figures, and speaker notes that actually say something ("Braga brings volume, Lisboa brings value"). One fact from the memo is missing - the 20 customers appear nowhere - and the title notes call it a "10-minute review", which the memo does not say.
- **Gamma.** 2026-09-18, the "Create with Agent" composer, free plan, 32 of the 400 signup credits. Four minutes, slide by slide, with its own illustrations - a shelf of wine and bowls on the title, a person at a laptop on the problems slide - a bar chart, a donut, tiles, cards and a closing. Export → PowerPoint gives a file in which every chart and illustration is a picture: the five region counts live only inside a chart image, the region revenues are rounded there to 948 and 791, and PowerPoint cannot edit any of it. The speaker notes are grey callout boxes on the face of every slide; the notes pane is empty. It added a monthly average (€1,131.83), category percentages (41.2 / 32.68 / 26.12) and "Ceramic Bowl - top Home product", and dropped the 20 customers and the €28.30 average. A "Made with Gamma" badge sits on every slide until you pay.
- **Canva.** 2026-09-18, free plan. There is no Canva AI on it: canva.com/ai is a 404, the home box searches templates, the presentation editor's tools and apps have no generator. Magic Design's prompt box is capped at 280 characters - the brief alone is 395 - and with the full text forced in it opened an Instagram-post design and returned "problems processing your request". Canva Docs' Convert menu, the notes-to-deck route, ends in a "Go to Canva Pro" button. Not run; the pricing page puts Canva AI 2.0 under Pro and gives Free "up to 20 AI uses".

## Deviations

ChatGPT's refusal (daily data-analysis allowance) stands: retried after the reset the next morning, refused again, because the Excel retry ran first and the free allowance covers one attempt a day. Kimi refused five times over six hours and answered on the sixth attempt; it is scored on that answer, with the free-tier rating carrying the wait. Readability was to be checked from a Keynote PDF export; Keynote hung on this machine, so the look column comes from the in-app previews (Claude, Gemini), one QuickLook thumbnail (Claude's title slide) and the structural check of every file. Gemini's answer is a canvas, not a file; it is scored on the PPTX its own export produced. Gamma and Canva, the Roster's two deck-builder Specialists, were run a day after the nine generalists, on 2026-09-18; Gamma is scored on its exported PPTX, Canva as unable to take the brief on the free plan. Gamma's facts count is the text the file carries; five region counts that are exact but only inside a picture are noted, not counted.