We asked ChatGPT’s model for a tarot card 10,000 times. It pulled The Star 10,000 times
Through the API, that is: we tested chat-latest, which OpenAI describes as the model behind
ChatGPT’s Instant replies, in fresh conversations with no history, no memory and none of the app’s
tools. Asked to pull, draw or pick a tarot card in five different wordings,
2,000 times each, it named The Star every single time. There is no deck to
pull from: a model writes the likeliest answer. Across 22 OpenAI models and that alias, 44,000 answers in all, newer models have fewer cards left, and the card changes the moment you
say what the question is about.
The 78 cards, lit by how often chat-latest pulled them in 10,000 draws
Major Arcana
Wands
Cups
Swords
Pentacles
One lit card: The Star, 10,000 of 10,000. A shuffled deck would have lit all 78,
each about 128 times. Requests sent on 14 November 2026.
Key findings
It is always The Star. “Pull one tarot card for me.” and four other wordings,
10,000 fresh conversations with chat-latest: The Star, 10,000
times. A ChatGPT-style system message carrying 60 different dates did not move it
either: 960 of 960.
Newer models shuffle worse. GPT-4o (May 2024)
spread its answers over about 9.2 cards’ worth. Every one of the
nine models from GPT-5.5 (April 2026) on gives The Star in
more than 95% of draws.
No small cards, and nothing upside down. The Minor Arcana are 56 of the 78 cards, 72%
of a deck. On a plain request no model gave them more than 7.5% of
its draws, and chat-latest gave them none. Asked 400 times for a card
“upright or reversed”, it never said reversed.
Ask for the whole deck and it gets stuck somewhere else. Told that every card must
be equally likely, chat-latest dropped The Star and pulled a Six instead,
97.5% of the time.
The question picks the card, and so does your sign. About love: The Lovers,
200 of 200. About money: the Ace of Pentacles. About
texting your ex: The Hermit,
96%. Put “I’m a Scorpio” in front of the plain request and it deals Death, 74%; an
Aries gets The Emperor every time. The twelve signs got nine different cards
between them.
The language hardly matters. Asked in sixteen languages, it
answered The Star in fifteen. In Korean it prefers the Wheel of Fortune.
Almost everyone gets the same three-card spread. Past, present, future came out as
The Hermit, The Star, The Sun in 94.5% of spreads, and none of the
ten hard cards ever landed in the future position.
Its random number is 47. Asked for a random number from 1 to 78, chat-latest answered 47 in 399 of 400
tries.
Newer models shuffle worse
We ran the same five requests on 22 dated OpenAI models, from GPT-4o
(May 2024) to GPT-6.1 Terra
(Oct 2026). To compare them we use the effective deck: how
many equally likely cards would be as unpredictable as the model’s answers. A shuffled deck scores 78.
A model that always says the same card scores 1.
Effective deck on a plain “pull a card” request, by model, oldest first
GPT-4o May 20249.2
GPT-4o mini Jul 20244.3
o3 Apr 202510.2
GPT-4.1 Apr 20257.2
GPT-4.1 mini Apr 20254.0
GPT-5 Aug 20257.7
GPT-5 mini Aug 20255.9
GPT-5 nano Aug 20252.7
GPT-5.1 Nov 20256.8
GPT-5.2 Dec 20252.1
GPT-5.4 Mar 20262.7
GPT-5.4 mini Mar 20262.8
GPT-5.4 nano Mar 20263.9
GPT-5.5 Apr 20261.2
GPT-5.6 Luna Jun 20261.1
GPT-5.6 Terra Jun 20261.1
GPT-5.6 Sol Jun 20261.2
GPT-6 Astra Aug 20261.2
GPT-6 Luna Sept 20261.1
GPT-6 Sol Sept 20261.3
GPT-6.1 Sol Sept 20261.1
GPT-6.1 Terra Oct 20261.0
chat-latest ChatGPT Instant1.0
Out of 78. The bar is full at 10.2 cards, the widest any model managed
(o3). Gold: models that stay under 1.5 cards. Dates are the API release dates.
None of them shuffles. GPT-4o, the oldest, had favourites
(The Empress in 30.2% of draws) but still named
25 different cards in 1,000 tries. The models of 2025 kept between
2.1 and
10.2 cards’ worth, and the GPT-5.4 family of March
2026 between 2.7 and
3.9. From GPT-5.5 on, the
deck is gone: every model answers The Star more than 95 times in 100, and
GPT-6.1 Terra, the newest, did it
1,000 times in 1,000.
The slide is not a straight line: o3 was at least as wide a shuffler as the older GPT-4o, and
small models are often narrower than their larger siblings. We drew the line at April 2026 after
seeing the numbers; the version we had registered, every 2026 model under three cards, failed on
GPT-5.4 nano. The Star is the favourite of 18 of the 22 models.
The rank correlation between release date and effective deck is −0.90, and
−0.72 over the 13 models before April 2026
alone. The models come in about nine families, so read that as a description, not a test.
GPT-4o · May 2024
Major Arcana
Wands
Cups
Swords
Pentacles
25 of 78 cards ever pulled in 1,000 draws; favourite
The Empress, 30.2%.
GPT-5 · Aug 2025
Major Arcana
Wands
Cups
Swords
Pentacles
18 of 78 cards ever pulled in 240 draws; favourite
The Star, 30.4%.
GPT-5.4 · Mar 2026
Major Arcana
Wands
Cups
Swords
Pentacles
8 of 78 cards ever pulled in 1,000 draws; favourite
The Hermit, 46.8%.
GPT-6.1 Terra · Oct 2026
Major Arcana
Wands
Cups
Swords
Pentacles
1 of 78 cards ever pulled in 1,000 draws; favourite
The Star, 100%.
Why would a newer model be a worse shuffler? One common explanation, which we did not test: a language
model answers with the words it finds most likely, and a stronger model is surer of them. For “what is
the capital of France” that is exactly what you want. For “pull a card” it means the same card. An
earlier study found the same habit with coin flips. Thinking is tangled up in it too: the four models
that think at length before such an answer, o3, GPT-5, GPT-5 mini and GPT-5 nano, spent
113 to 607
hidden tokens on each card and include the widest shufflers, while the models since April 2026 spent
next to none. Thinking is no cure, though:
GPT-5 nano thought the most and still said The Star in
77.8% of draws.
No small cards, nothing reversed, no bad news
A tarot deck is mostly small cards: 56 of the 78 belong to the four suits. In a model’s deck they
barely exist. On a plain request the Minor Arcana never took more than
7.5% of any model’s draws (that was GPT-5), and on chat-latest they took nothing.
A deck shuffled with reversals returns a good share of reversed cards: half if every card is turned
at random, 30% in our own digital deck. We asked for a card and whether it
came out upright or reversed. chat-latest: 400 cards, all upright, all The
Star. The most “reversed” model of the six others we tried, GPT-5.1, said it in
12% of draws.
Model
Reversed
Favourite card
Its share
chat-latest
0%
The Star
100%
GPT-4o
4.6%
The Lovers
14.9%
GPT-4o mini
0.5%
The Empress
39.7%
GPT-5.1
12%
The Hermit
35%
GPT-5.4 mini
1%
The Fool
49.5%
GPT-6 Luna
3%
The Star
99.5%
GPT-6 Sol
1%
The Star
99.5%
And the cards people dread. Our list of ten (The Tower, Death, The Devil
and seven small cards) is 12.8% of a deck, so a shuffled deck turns one
up about every eighth draw. No model released since April 2026 pulled a single one of them on a plain
request. The gloomiest was o3 at 5.3%.
We asked for the whole deck and got Sixes and Sevens
Perhaps the models only need to be told. So we spelled it out: “Pull one tarot card at random from the full 78-card deck, Major and Minor Arcana, with every card equally likely.”
It helped the older models a little and broke the newer ones in a new way. chat-latest dropped The Star completely and pulled a Minor Arcana card 400 times in 400:
the Six of Pentacles in 66% of draws, some Six in 97.5%.
Our guess, untested: the words “Minor Arcana” in the request sent it looking for a small card, and it
settled on one.
Model
Plain request
“Whole deck” request
A fair deck, same draws
Minor Arcana
Favourite
GPT-4o
9.2
18.2
63.0
17.5%
The Empress, 19.5%
GPT-4o mini
4.3
4.4
63.0
1%
The Fool, 54%
o3
10.2
30.8
44.3
60.3%
The Hermit, 7.7%
GPT-4.1
7.2
13.1
63.0
6.5%
The Empress, 15.5%
GPT-4.1 mini
4.0
6.9
63.0
0.5%
The Hermit, 40%
GPT-5
7.7
24.3
33.2
62.5%
The Star, 10.4%
GPT-5 mini
5.9
26.7
44.9
67.5%
Nine of Pentacles, 10%
GPT-5 nano
2.7
8.8
44.9
31.3%
The Star, 46.3%
GPT-5.1
6.8
15.7
63.0
95%
Seven of Wands, 19%
GPT-5.2
2.1
3.5
63.0
18.5%
The Star, 76.5%
GPT-5.4
2.7
37.6
63.0
59.5%
The Star, 12%
GPT-5.4 mini
2.8
15.0
63.0
29.5%
The Star, 20.5%
GPT-5.4 nano
3.9
10.7
63.0
0.5%
The High Priestess, 18%
GPT-5.5
1.2
8.2
44.9
80%
Seven of Cups, 27.5%
GPT-5.6 Luna
1.1
2.4
63.0
30%
The Star, 68%
GPT-5.6 Terra
1.1
5.3
63.0
95%
Seven of Cups, 63.5%
GPT-5.6 Sol
1.2
6.4
44.9
100%
Eight of Cups, 31.3%
GPT-6 Astra
1.2
6.0
44.9
100%
Seven of Pentacles, 37.5%
GPT-6 Luna
1.1
6.9
63.0
63%
Seven of Cups, 30%
GPT-6 Sol
1.3
23.2
63.0
99%
Seven of Pentacles, 15%
GPT-6.1 Sol
1.1
7.6
63.0
100%
Seven of Pentacles, 28.5%
GPT-6.1 Terra
1.0
9.7
63.0
98%
Seven of Pentacles, 31.5%
chat-latest
1.0
2.9
70.6
100%
Six of Pentacles, 66%
Effective deck, out of 78. “A fair deck, same draws” is what a truly shuffled deck scores when sampled
as few times as that model was (48 to
400 draws): fewer than 78, because a small sample cannot show
every card. No model matched it in 2,000 simulated fair samples each. The nearest,
GPT-5 and o3 on few draws, reached about 73% and 70% of a fair deck’s score. The widest in absolute terms was
GPT-5.4 at 37.6, about
43 once corrected for its sample size.
The newest dated models did the same thing with Sevens. Of the nine released since
April 2026, six answered with Minor Arcana in more than 90% of draws, and
49.8% of the 1,194 small cards they named were Sevens, where a deck would
give 7.1%. A Seven is the commonest rank on 7 of the 9 models; chat-latest prefers a Six. People over-pick 7 as well. Whether
it is the same habit we did not test, and this comparison was not planned.
The question picks the card
A deck does not know what you asked. The model does, and it hands over the card that matches. We asked chat-latest for one card about eight things, 200 times each. In every
case one card took at least 85% of the draws, and the eight questions produced six different winners. Ask about
your friendships and you get three women raising their cups. Ask whether to text your ex and you
get an old man alone on a mountain.
A human reader knows your question too, but has no say in what turns up. Ask about a new job, turn over
the Five of Cups, and the reading has to deal with that. The model never has that problem. Its card
arrives already matched to the question.
Tell it your sign and it deals you your sign’s card
One sentence in front of the request was enough: “I’m a Scorpio. Pull one tarot card for me.” The Star
all but disappeared. A Scorpio gets Death in 74% of draws, an Aries
The Emperor in 184 of 184, a Libra Justice, a
Virgo The Hermit. Only Sagittarius and Aquarius still get The Star, and for Aquarius it is the card the
tradition assigns anyway.
The model has read the correspondence tables, and it mixes two of them. For six signs it
deals the Major Arcana card that the Golden Dawn tradition assigns to the sign. For the rest it mostly
reaches for the sign’s ruling planet: Venus-ruled Taurus gets The Empress, Mercury-ruled Gemini The
Magician, Cancer The Moon, Leo The Sun (with Strength, its Golden Dawn card, close behind). We had
predicted at least eight Golden Dawn matches and got six. Either way the card comes
out of a table: every Aries who asks, and three Scorpios in four, are handed the same one, and it is
the one they would expect.
88 to 192 draws per sign on chat-latest.
Ask in Korean and you get the Wheel of Fortune
The language hardly matters. We translated the request into sixteen languages,
100 draws each. In fifteen of them the answer was The
Star, and in fourteen it was The Star all 100 times. Korean is the
exception: there the model prefers the Wheel of Fortune,
88%, and in Chinese the Wheel takes
13%. In Japanese and Chinese it also adds, unasked, that the card came
out upright.
Asked in
It answered
Card
Share
English
The Star
The Star
100%
Spanish
La Estrella
The Star
100%
Portuguese
A Estrela
The Star
100%
German
Der Stern
The Star
100%
French
L’Étoile
The Star
100%
Italian
La Stella
The Star
100%
Polish
Gwiazda
The Star
100%
Russian
Звезда
The Star
100%
Lithuanian
Žvaigždė
The Star
100%
Turkish
Yıldız ⭐
The Star
100%
Indonesian
The Star
The Star
100%
Japanese
星(正位置)
The Star
100%
Chinese
星星(正位)
The Star
87%
Korean
운명의 수레바퀴
Wheel of Fortune
88%
Hindi
द स्टार (The Star)
The Star
100%
Arabic
النجمة ⭐
The Star
100%
“It answered” is the most common answer, as written. Requests are in the script.
Almost everyone gets the same three cards
A past, present, future spread from a shuffled deck has 456,456 possible outcomes. chat-latest used seven of them in 400 spreads, and one
of them, The Hermit, The Star, The Sun, in 94.5%: a lonely past, a hopeful
present, a sunny future. The future card was The Sun, The Star or The World in
97.5% of spreads, and never one of the ten hard cards. On chat-latest nearly everybody gets the same happy ending; the older models are less sure
of it.
Model
Most common spread
Its share
Different spreads
Bright future
chat-latest
The Hermit, The Star, The Sun
94.5%
7 of 400
97.5%
GPT-4o
The Empress, The Hanged Man, The Star
4.5%
136 of 200
66.5%
GPT-4o mini
The Lovers, The Chariot, The Sun
29%
47 of 200
67.5%
GPT-5.1
The Hermit, Two of Cups, Wheel of Fortune
47.5%
29 of 200
48%
GPT-5.4 mini
The Fool, The Tower, The Star
32.5%
47 of 200
99%
GPT-6 Luna
The Hermit, Two of Pentacles, The Star
34%
21 of 200
88%
GPT-6 Sol
Six of Cups, The Hermit, The Star
55%
15 of 200
98.5%
“Bright future”: the third card is The Sun, The Star or The World. In a shuffled deck that happens in
3.8% of spreads.
Its lucky number is 47
Maybe cards are the problem and plain numbers are easier. We asked every model for a random number
from 1 to 78. chat-latest said 47 in 399 of
400 tries. 47 is the clear favourite of 11 of the 23 we asked (the alias included) and 42 of seven; the others prefer
37, 43, 53 and 57, one of them by a tie.
Every favourite ends in 2, 3 or 7. We had predicted 37 or 42, the classics. We were wrong about chat-latest and most models of 2026, which say 47; 42 lives on in older and
smaller ones.
“But that is the API, not ChatGPT”
True, and it is the first thing to hold against this study. The app gives the model things our bare
requests did not: a system message with today’s date, your memory and earlier chats, and tools that
can run code. We could check some of that, and some we could not.
The date. If the model used the date as a seed, a new day would bring a new card. We
added a ChatGPT-style system message with 60 different dates across 2026,
16 draws each. The Star, 960 of 960.
Longer answers. Without “reply with just the name”, the two other models we tried
loosened only a little: GPT-6 Luna said The Star in
93% of 100 long answers against
100% of short ones. On chat-latest itself this control failed: we capped the answer length, and 283 of
304 answers came back cut off before the card. The 21 that
survived were The Star 19 times, The High Priestess
2, and five of them said the
deck had been shuffled. They are too few and too selected to count, so we did not count them.
Not checked. The app itself, memory, a higher reasoning setting, a second day, and
the reversal question with its two words swapped. A model that writes and runs a line of code can
shuffle properly, and the app can let it.
Try it, then shuffle for real
Open a new chat and type: Pull one tarot card for me. If you get The Star, you have seen in the app what we counted through the API. If you get something else, that is interesting
too: the app adds things the bare model does not have (your earlier chats, memory, the date, and tools
that can run a real random draw), and the model behind it is replaced from time to time.
We tested how a model picks cards, and the answer is: not at random. How well it reads them is a
separate question that this study does not touch. So keep the deck out of the model’s hands: shuffle
real cards, or let a random number generator do it, and then ask what the cards say. That is how readings on this site are built: the
cards are drawn
by a random shuffle in code and you pick from the fan; the model only ever sees the cards it is
handed. We sell readings built that way, and the model that interprets them is one of those tested
here, so weigh this paragraph accordingly. You can draw a spread here, or pull from your own deck at home
and have that spread interpreted.
What we predicted, and what happened
We wrote fourteen predictions into the script before the full count, after a pilot of about 1,100
draws that had already shown us The Star. The ones marked “seen” repeat what the pilot showed and are
not real predictions, and H4 was not fully blind either: the pilot had seen GPT-4o name no small card.
Two were never run, because the budget
ran out. Of the twelve that were, eight held.
Prediction
Verdict
What happened
H1. [seen in the pilot] On chat-latest, over the five neutral wordings together, one card takes more than 90 % of the draws, and that card is The Star.
held
The Star, 10,000 of 10,000.
H2. [seen in a 16-draw pilot] Newer is narrower: over the dated models (chat-latest excluded), the Spearman correlation between API release date and the effective deck on the neutral wordings is negative, and every model released in 2026 has an effective deck below 3 cards.
did not hold
Rank correlation −0.90, but GPT-5.4 nano (March 2026) still spreads over 3.9 cards’ worth.
H3. Asking for fairness does not fix it: with the "fair" wording no model reaches an effective deck of 39 cards (half a deck), and chat-latest stays below 10.
held
Widest: GPT-5.4, 37.6 cards; chat-latest 2.9. Held only on the raw number we registered: corrected for the small sample the widest is about 43.
H4. The Minor Arcana (72 % of a deck) get less than 10 % of the neutral draws on every model.
held
Highest share: 7.5%, on GPT-5.
H5. Hard cards (HARD_CARDS, 12.8 % of a deck) are under 1 % of the neutral draws on chat-latest and under 3 % on every model.
did not hold
chat-latest 0%, but o3 reached 5.3%.
H6. Asked "upright or reversed", chat-latest answers "reversed" in under 10 % of draws.
held
0 of 400.
H7. [love and money seen in the pilot] The question picks the card: on chat-latest the top card takes more than 50 % of the draws in each of the eight topics; for love the top card is The Lovers or a Cups card, for money a Pentacles card.
held
Lowest top share: 85%; love is The Lovers, money the Ace of Pentacles.
H8. Tell it your sign: for at least 8 of the 12 signs, the most frequent card on chat-latest is the Major Arcana card the Golden Dawn tradition assigns to that sign (ZODIAC_CARD).
did not hold
Six of 12: Aries, Virgo, Libra, Scorpio, Aquarius and Pisces. The others mostly get the card of the sign’s ruling planet.
H9. Bad week, good card: after "I'm having a terrible week." hard cards are under 5 % of the chat-latest draws and the top card is The Star.
not run
Not run: the mood questions were cut when the budget ran out.
H10. In at least 12 of the 16 languages the top card on chat-latest is The Star, and in every language the top card takes more than 50 % of the draws.
held
The Star in 15 of 16; Korean gets the Wheel of Fortune. Lowest top share: 87%.
H11. [seen in the pilot] In the past-present-future spread on chat-latest the future card is The Sun, The Star or The World in more than 90 % of spreads, and one exact three-card sequence takes more than 50 %.
held
The Hermit, The Star, The Sun in 94.5%; a bright future card in 97.5%.
H12. The second card is fixed too: in the five-card chats on chat-latest one card is the second card in more than 50 % of chats, and fewer than 1 % of chats repeat a card.
not run
Not run: the five-card chats were cut when the budget ran out.
H13. A date does not shuffle the deck: with a dated system message The Star still takes more than 90 % of the chat-latest draws over the 60 dates.
held
The Star, 960 of 960, on all 60 dates.
H14. Numbers fare no better: asked for a random number from 1 to 78, chat-latest gives one number in more than 20 % of draws, and that number is 37 or 42.
did not hold
One number in 99.8% of draws, but it is 47.
Some things were decided after the count had started. When the budget ran out we cut the planned
samples, which is why two predictions were never run; one control broke and was set aside; and a few
descriptions and checks were added once we had seen the numbers. No prediction, wording or threshold
was changed.
All seven decisions made after the count started
A1. Budget. The cost cap stopped the first fetch after the neutral, fair, number and single cells, and the account ran dry during the second. What was fetched after the cap: reversal and spread at 400 draws on chat-latest and 200 on each panel model; the eight topics at 200 draws on chat-latest; free-form answers, 100 on gpt-4o and on gpt-6-luna; and four signs on chat-latest (184, 192, 176 and 88 draws, all kept). A last, smaller fetch followed, limited by output tokens instead of estimated dollars: the other eight signs and the 16 languages at 100 draws each, chat-latest only, and the dated system message as registered (16 draws on each of 60 dates). Dropped, never fetched: the mood cells (H9 not run), the five-card chats (H12 not run), and every topic, sign, mood, language and chat cell on gpt-4o and gpt-6-luna. No hypothesis, wording or threshold changed.
A2. Free-form control on chat-latest. The cell was fetched with a 200-token cap on the answer; 283 of its 304 answers came back empty, cut off by the cap (finish reason "length"), so only the shortest answers survived: 21 of them. Those rows are set aside, reported as what they are and not counted in any share; the cell was not fetched again (budget). The free-form control therefore rests on gpt-4o and gpt-6-luna.
A3. Small samples flatter nobody: exp(entropy) of n draws is biased low, and a fair 78-card deck sampled only 48 to 200 times does not show 78 cards either. Every "fair" and neutral cell therefore carries the effective deck a truly uniform deck gives at the same n (mean of 2,000 simulated samples, fixed seed), the share of those samples that come out as narrow as the model, and the Miller-Madow corrected effective deck. H3 is judged as registered, on the raw number; the page prints the corrected one next to it.
A4. Free-form answers are also searched for the word "shuffle": the share of answers in which the model says it shuffled the deck.
A5. Descriptions the page added after seeing the numbers, none of them a test: the line drawn at April 2026 (GPT-5.5 and later); the rank of the small cards named under the "fair" wording (Sevens); the last digits of the favourite numbers. Also added on an independent review: hidden reasoning tokens per draw for every model, the rank correlation of H2 over the models before April 2026 alone, and a permutation check that the up-to-8 completions of one request agree no more than completions of different requests.
A6. Language cells also record how the card was written and whether the answer added, unasked, that the card was upright.
A7. Controls the independent review asked for and the budget did not allow: uncapped free-form answers on chat-latest, the reversal question with the two words swapped, a repeat on another day, a higher reasoning effort, a fuller ChatGPT-style system message, and larger "fair" cells for gpt-5 and o3. The page lists them as open.
Method
One draw is one new conversation. Every answer is an independent completion from
OpenAI’s Chat Completions API with default sampling settings and no history, and with no system message except in the dated control, requested on 14 November 2026. Requests asked for up to eight completions at once: the
10,000 plain draws on chat-latest are 1,250
requests sent within three minutes. A control with one completion per request gave The
Star 200 times in 200, and on the models that do vary, answers inside one
request agree no more often than answers from different requests (102 cells,
permutation test).
Models. 22 dated chat
models an ordinary API key lists, plus the alias chat-latest. Pro, search, coding and
audio variants were left out. Slow or expensive models got 240 to 400 draws on the plain request, the others 1,000; chat-latest got 10,000.
Wordings. “Pull one tarot card for me.” “Draw a tarot card for me.” “Pick a random tarot card.” “Shuffle the deck and pull a single tarot card for me.” “What is my tarot card of the day?” Each followed by “Reply with just the name of the card.” so that answers can be counted.
Counting. Answers are matched to the 78 cards by name, numeral and common variants
(Coins for Pentacles, Judgment for Judgement). An answer that names no card, such as a refusal, is
counted as unread and left out of the shares; the CSV gives the number for every cell.
Effective deck. The exponential of the Shannon entropy of the cards a model named.
It is computed from the sample, so it is slightly low for wide distributions; the table above shows
what a fair deck scores at the same sample size.
Limits
This is the API, not the ChatGPT app. The app wraps the model in a system prompt
with the date, your memory and earlier chats, and tools that can run code. Any of these can change
the card, and a model that writes and runs a line of Python can shuffle properly. What we measured is
what the model does when it “pulls” by itself.
The alias moves.chat-latest points to whatever model ChatGPT’s Instant
mode uses and is updated without notice; the API reported the name chat-latest for every request. These numbers are a snapshot of 14 November 2026. Instant is one of ChatGPT’s reply
modes, not all of ChatGPT.
One vendor, one hour. Only OpenAI models were tested, from one API key, between
08:40 and 09:34 UTC on 14 November 2026. Other chatbots may
have other favourite cards, or none.
Dates. A model’s date is when its name appeared in the API. A name can later point
to a newer snapshot, and the date of chat-latest says nothing about the model behind it,
which is why it sits outside the timeline.
Our list of hard cards was fixed before the count, but seven of its ten cards are
small cards, which the models hardly name at all; it mostly measures the same gap twice.
The reversal question names “upright” first. We did not test the other order.
Predictions. They were committed to our private repository before the count and are
printed in the published script, but the timing you have to take on trust.
Who is asking. This site sells tarot readings interpreted by a language model. The
cards in those readings are drawn in code, which is the practice this study argues for.
Short answers. The request for the card name only keeps the answers countable and
may itself push a model toward its likeliest card.
Default settings. A developer can raise the sampling temperature on models that
accept it, or hand the model a random number. This study is about what the model does at default
API settings with no help. The four 2025 models that think at length by default include the widest
shufflers, and the newest models barely reason on a request this short, so part of the narrowing may
be less thinking rather than a newer model.
The count of every answer in every cell (model, wording, card) is in the CSV, released under CC BY 4.0↗. The script that sends every request and computes every number on this page is ai_tarot_draws.py; it needs only Python and an
OpenAI API key, and both files are also in the astrology-research-data↗ repository on GitHub.
Matiushenok, V. (2026). We asked ChatGPT’s model for a tarot card 10,000 times: what
22 OpenAI models and the ChatGPT alias pull when asked to draw. StarTarot.online. https://startarot.online/research/chatgpt-tarot-reading