Skip to content

BlogAI + Anki

ChatGPT vs Claude vs Gemini for Anki cards: I tested all 3

A short handout and a dense 22-page chapter with 15 diagrams, three runs each, every card graded blind. Here's where each one breaks.

Omar··9 min read


On a short handout, all three chatbots write good cards. On a dense chapter with diagrams, AnkiDecks wins clearly: 99% correct, 1 repeat instead of 7–22, and 109 diagram cards instead of 0. Claude is the best chatbot.

Lots of students make Anki cards with a chatbot now. I wanted to know which one is actually best at it, so I tested them instead of guessing. Twice: once on an easy file, once on a hard one.

The results at a glance

Wrong or made-up cards

Lower is better

  • ChatGPT
  • Claude
  • Gemini
  • AnkiDecks

Repeated cards per deck

Lower is better

  • ChatGPT
  • Claude
  • Gemini
  • AnkiDecks

Diagram cards

Image occlusion · higher is better

  • ChatGPT
  • Claude
  • Gemini
  • AnkiDecks
Test 2: a 22-page textbook chapter with 15 diagrams. Errors from a random 80 cards per deck, graded by 2 judges; repeats averaged over 3 runs. Every number is in the tables below.

How I tested it

  • Test 1, easy: an 11-page cardiology handout from the NIH’s free training notes. It’s our sample PDF, so you can repeat it.
  • Test 2, hard: a 22-page textbook chapter on heart anatomy (about 9,000 words) with 15 labelled diagrams, from OpenStax Anatomy and Physiology, a free textbook.
  • The models behind the apps: GPT-5.6 Luna (free ChatGPT’s default since August, reported by GIGAZINE), Claude Sonnet 5.5 and Gemini 3.8 Flash. I ran them through their APIs with the same page text AnkiDecks read. The apps read PDFs themselves, so they can behave a little differently.
  • One prompt: the medical prompt from our free prompt page, asking for exactly as many cards as AnkiDecks made from the same file (24, then 276), so every deck is the same size.
  • Three runs each, because the same chatbot gives different cards every time.
  • Graded blind. Two AI judges (Claude Opus 5.5 and GPT-5.6 Sol) checked cards against the text without knowing which tool wrote them: every card in test 1, a random 80 per deck in test 2. Repeats were counted on every card: two cards with almost the same wording.

Can ChatGPT make Anki cards? Test 1: a short handout

Yes, with one catch. First I typed what most people type: “Make Anki flashcards from this lecture.” All three said yes, and each answered in a different shape.

Below are Anki-ready flashcards in **tab-separated format**. Copy the contents into a `.tsv` or `.txt` file and import into Anki using **Tab** as the field separator.

```tsv
What type of organ is the heart?	The heart is a four-chambered muscular organ.
What is the approximate size and shape of the heart?	It is roughly the size and shape of a man's closed fist.
Where is most of the heart's mass located relative to the midline?	Approximately two-thirds of the heart's mass lies to the left of the midline.
What encloses the heart?	The heart is enclosed in a pericardial sac.
What lines the pericardial sac?	The pericardial sac is lined by the parietal layer of a serous membrane.
What is the visceral layer of the serous pericardium called?	The epicardium.
What are the three layers of the heart wall?	Epicardium, myocardium, and endocardium.
What is the outermost layer of the heart wall?	The epicardium.
What is the middle layer of the heart wall?	The myocardium.
What is the innermost layer of the heart wall?	The endocardium.

55 cards in a code blockGPT-5.6 Luna

The real start of each answer to “Make Anki flashcards from this lecture”. Tap a name to switch.

ChatGPT and Claude gave text you can save as a file for Anki. Gemini gave formatted cards that Anki can’t import, so you’d copy them by hand. Claude also added a fair note that one sentence on page 11 is oddly worded. So the catch: tell the chatbot the exact format. With our prompt, all three answered in Anki-ready lines:

How many chambers does the human heart have?;Four
What forms the epicardium?;The visceral layer of the serous membrane
What are the three layers of the heart wall?;Epicardium myocardium and endocardium
Which chambers receive blood from veins?;The atria
Which chambers forcefully pump blood out of the heart?;The ventricles
What blood does the right atrium receive?;Deoxygenated blood from systemic veins
What blood does the left atrium receive?;Oxygenated blood from pulmonary veins
What are valves between atria and ventricles called?;Atrioventricular or cuspid valves

Asked for 24 · got 25GPT-5.6 Luna

Our prompt, first run: one card per line as front;back.
Short handoutChatGPTClaudeGeminiAnkiDecks
Fully correct97%99%95%100%
Wrong, or added a fact0%0%4%0%
Made the 24 cards asked0 of 3 runs3 of 33 of 324
Page on every cardNoNoNoYes
Test 1: 72 cards per chatbot (3 runs) and 24 from AnkiDecks, each graded by 2 judges. Swipe sideways to see every column.

Honestly, closer than I expected. On a short, clean handout all three write good cards, and a chatbot is free and done in seconds. The one thing none of them do is tell you where each card came from:

Basic · page 5

When the ventricles contract, which valves close and why?

Atrioventricular valves close to prevent backflow into the atria.

From your lecture, page 5When the ventricles contract, atrioventricular valves close to prevent blood from flowing back into the atria.
A card from the AnkiDecks deck for test 1. Checking it takes a glance at the quote.

Which AI is best for a long PDF? Test 2: a dense chapter

Real lectures aren’t 10 clean pages. So I gave them 22 dense textbook pages and asked for 276 cards. This is where they start to slip.

Dense chapterChatGPTClaudeGeminiAnkiDecks
Fully correct95%98%90%99%
Wrong0.6%0%3.5%0%
Added a fact not in the chapter0.8%0%7.3%0.6%
Repeated cards per deck7–168–1414–221
Made the 276 cards asked2 of 3 runs2 of 31 of 3276
Image occlusion cards000109
Page on every cardNoNoNoYes
Test 2: a random 80 cards per deck graded by 2 judges (480 grades per chatbot, 160 for AnkiDecks). Repeats counted on all cards. Swipe sideways to see every column.
  • Gemini got worse with length: 3.5% wrong, and 7.3% added things the chapter doesn’t say. Some are true, but on an exam you’re marked on your course, not the internet. It also made 360 and 338 cards when I asked for 276.
  • All three repeat themselves: 7 to 22 near-identical cards per deck.
  • Claude stayed accurate (98%, nothing made up). It’s the one I’d pick if you use a chatbot.

What the repeats cost you

A repeat looks harmless, but Anki shows it to you again and again. Using the model behind our review calculator, one card costs about 7.5 reviews over a 4-month semester at Anki’s default 90% retention. So per chapter like this:

  • ChatGPT and Claude (about 10 repeats): about 75 wasted reviews, roughly 12 minutes.
  • Gemini (about 18 repeats): about 135 wasted reviews, roughly 22 minutes.
  • AnkiDecks (1 repeat): about 8 reviews, roughly 1 minute.

Multiply that by every lecture in a semester. AnkiDecks writes more cards than it keeps: for this chapter it drafted 374 and threw out 98 before I saw them (43 low-value, 24 weak fill-in-the-blanks, 23 that asked two things at once, 6 repeats, 1 whose quote wasn’t in the chapter and 1 that pointed at “the slide”).

Can ChatGPT make image occlusion cards?

No, not in my test. Anatomy is half diagrams, and in Anki you learn them with image occlusion: boxes hide the labels, and you name what’s under each one. I gave each chatbot four of the chapter’s diagrams and asked for image occlusion cards.

I can make these as **image-occlusion cards with one card per label**. The four diagrams contain approximately **70 occlusions**:

1. **Dual circulation diagram** — 28 labels  
2. **Internal structures of the heart** — 25 labels  
3. **Chordae tendineae and papillary muscles photograph** — 3 labels  
4. **Coronary circulation, anterior and posterior views** — 24 label instances  

I’ll use the standard format:

- **Front:** original diagram with one label and its leader line occluded  
- **Back:** same diagram with the hidden label revealed  
- Repeated labels in separate locations will receive separate cards.
- The textbook paragraph text will be excluded.

The labels covered are:

### 1. Dual circulation

A list of labels · 0 cardsGPT-5.6 Luna

The real start of each answer when asked for image occlusion cards from four diagrams.
  • ChatGPT listed about 70 labels, then asked me to upload each diagram cropped. (The ChatGPT app can run code, so it might get further there. Through the API it made no cards.)
  • Claude said straight away it can’t make image files, and gave a 7-step guide to do it by hand in Anki, “about 5 minutes per figure”.
  • Gemini gave a checklist and the same by-hand steps.

At 5 minutes a figure, the chapter’s 12 usable diagrams are an hour of drawing boxes. AnkiDecks found them, boxed 182 labels and made 109 cards from the key ones:

A heart diagram (coronary circulation, front view) with key labels covered by boxes, and one box marked with a question mark: the label to name.
Part of an AnkiDecks image occlusion card from the chapter: name the label under the question mark. Diagram: OpenStax Anatomy and Physiology 2e, CC BY 4.0.

What this means for your files

Your fileWhat I’d use
A short handout or your own notes (up to ~10 pages)Claude with our prompt, or AnkiDecks if you want the page on every card.
A long PDF or textbook chapterAnkiDecks. Fewer wrong cards, almost no repeats, and every card checked against its page.
Slides or PDFs with labelled diagramsAnkiDecks. The chatbots made 0 diagram cards; you’d box them by hand.
A lecture you’ll be examined onAnkiDecks. Checking a card takes a glance at its quote, not a search through the slides.

What each one is best for

  • Claude: the best chatbot. Most accurate in both tests and nothing made up. One slip in test 1: a semicolon inside a card, which splits into the wrong fields when you import it. If you like Claude, AnkiDecks works inside Claude too.
  • ChatGPT: fine, but doesn’t count. I asked for 24 cards and got 25, 27 and 26. It also dropped commas in lists (“Epicardium myocardium and endocardium”).
  • Gemini: good on short texts. On the long chapter it added the most outside facts and made the most repeats.
  • AnkiDecks: for lectures you’ll be tested on. Most accurate, 1 repeat, image occlusion from your diagrams, and the line and page on every card. It’s slower (7 minutes for this chapter in my test). The free plan covers 15 pages a file and 20 diagram cards a month; Pro covers 100 pages a file and 300 diagram cards a month, and Max has no monthly limit on diagram cards. See plans.

My pick: a chatbot for a quick handout, AnkiDecks for real lectures. That’s what I do, and it’s why I built it. Want to check my work? Here’s every answer, prompt and grade from both tests (zip).


Try it on your next lecture

Every card with its page, plus diagram cards. 3 free decks a month.

Start free

Free to try · No credit card needed

Questions

Which AI makes the best Anki cards?
Of the three chatbots, Claude: 98–99% fully correct in both tests and no made-up facts. On a dense chapter AnkiDecks did better still (99%, 1 repeat instead of 7–22), and it was the only tool that made image occlusion cards from diagrams.
Can ChatGPT make image occlusion cards for Anki?
Not in my test. Given four diagrams, ChatGPT, Claude and Gemini all listed the labels and explained how to draw the boxes in Anki yourself, but none made a card you could import. AnkiDecks made 109 from the chapter's 12 usable diagrams.
Do AI chatbots make up facts on flashcards?
Rarely on a short text. On the dense chapter, 7.3% of Gemini's graded cards added something that wasn't in the chapter and 3.5% were wrong. ChatGPT was under 1% and Claude 0%. None of them show the page a card came from, so you can't check them quickly.
Is it OK to use ChatGPT for Anki cards?
For a short handout or your own notes, yes: ask for one card per line in a fixed format and check every card. For long lectures, watch for repeats and facts from outside your course, and diagrams still have to be boxed by hand.
Is AnkiDecks just ChatGPT with extra steps?
No. It checks every card against your file and drops any card whose quote isn't in it, removes repeats, keeps the page and exact line on each card, turns diagrams into image occlusion cards, and downloads a ready Anki deck.