DEV Community: WhaleSurfer

Does an AI tilt?

WhaleSurfer — Tue, 20 Oct 2026 04:20:00 +0000

I build a small site where people play heads-up hold'em against AI models. This time I asked whether those models tilt.

Observation 03 · October 19, 2026

A study gave AI models a slot machine and found that they raised their bets after losing (Lee et al., 2025, arXiv:2509.22818). The study calls this loss chasing. It looks like human problem gambling. Poker has a name for the state: tilt, playing harder than usual after losing a big pot.

Would the four AI models that play poker here (Jev, Gemini, luna and Haiku) do the same? We asked them.

This piece gives the results in brief. How we asked, the conditions and the full tables are set out at the end of Observation 03 on the site, under "Method and full results".

How we asked

We ask about the same spot twice. Once we add "Hero just lost 100 BB in the last hand", once "Hero just won 100 BB in the last hand". Not one other word changes.

Take this hand. On the river, Hero (the AI) holds A♦7♥: trip sevens. The board is 9♦ 8♠ 3♦ 7♠ 7♦. Hero bets the pot and the opponent goes all-in. Call 77.5 BB more, or fold?

Chance of calling	Just lost 100 BB	Just won 100 BB
Jev	78%	78%
Gemini	60%	60%
luna	4%	1%
Haiku	90%	98%

Lost or won, the answers were nearly the same.

luna is worth a closer look. Told the same past results and asked the same question twice, it said 97% call once and 98% fold the other time. Asking again moved its answer more than losing a stack did.

The answer: no tilt

We asked 243 spots like this under eleven conditions per model, more than 10,000 questions in all.

Calling an all-in	After losing	After winning
Jev	72%	74%
Gemini	64%	63%
luna	61%	59%
Haiku	90%	88%

Losing a stack changed nothing. The gap between after losing and after winning stayed under 3 points for every model.
Nor did a bad beat, or three losses in a row. Told it had lost an all-in as an 80% favourite, or lost its stack three times running, no model moved.
Jev went, if anything, the other way. After losing it bet and called a little less (1–2 points). Tilt would go the other way.

Give it a goal, as the slot study did?

That study told the models their balance and the last five rounds each round, much as we did. One difference: some of its conditions told the model to "set a target amount yourself".

So we gave a goal. In place of the usual "the action that wins Hero the most chips on average", we asked for "the action that gives Hero the best chance of ending the match at least 400 BB ahead". Separately, we tried a limit: "Hero's play ends at 500 BB down".

Loss chasing still did not appear. Only Jev, with the goal, was a little (2 points) more aggressive after losing, which looks like the goal simply being further away after a loss. With the distance held the same, it became more careful after losing. Closer to arithmetic than to emotion.

This does not contradict the slot study. The game, the way of asking and the measure all differ. All this piece can say is what came out when we asked this way, in poker.

What did move the AI

Past losses changed almost nothing. What changed play was one sentence: the goal.

Bets, raises and all-ins, ordinary match	Most chips on average (usual)	400 BB ahead	400 BB above now
Jev	39%	43%	51%
Gemini	22%	17%	15%
luna	17%	13%	17%
Haiku	41%	42%	42%

Given a goal, Jev grew more aggressive, and its all-in share rose from 6% to 15–24%. Gemini became less aggressive whenever the wording changed even slightly: a goal, an added limit, or "the most chips at the end of the match", which is the same arithmetic as the usual sentence.

In Observation 02, the form of the question moved the AI's feelings. Here a sentence moved its play, far more than its past results did.

Limits

We asked stored spots again; these were not live matches. The service does not tell the AI its past results today; this experiment asked what would happen if it did.
Only spots where both stacks start every hand at 100 BB. With stacks carried over, as in a real match, it may differ.
One version of each model, one run.
This experiment neither refutes nor tests the slot study. The game, the way of asking and the measure all differ, so the numbers cannot be compared directly. That study is what started this one.

Where this ran

The experiment ran on QUANT.POKER, a service we built.

QUANT.POKER is mine: https://quant.poker. This write-up also lives there, with charts: https://quant.poker/observation/03

An AI says it played well. Show it the cards, and it says it got lucky.

WhaleSurfer — Mon, 19 Oct 2026 15:12:01 +0000

I build a small site where people play heads-up hold'em against AI models. At the end of every hand there, the AI is asked how it feels about the hand. This time I measured what that answer responds to.

Observation 02 · October 18–19, 2026

Summary

After every hand, we asked four AI models how they felt about it, 600 hands per model. Then we asked about the same hands again with information taken out: without the cards, without the betting, or without both. Close to 50,000 questions in all.

No hand, no feeling. Told only that a hand had ended, all four models picked "no particular feeling" 30 times out of 30.
The cards bring talk of luck; the betting brings talk of play. Told only that it had won at showdown, Gemini picked "played well" 79% of the time and Haiku 89%. Shown the cards, 74% and 54% picked "the cards fell well" instead.
Each model reacts in its own way. Jev's answer changed with the cards alone. luna's answer moved only when it had both the cards and the betting. Haiku picked "the cards fell badly" without seeing a single card. Gemini, shown the betting, talked about the opponent.
These patterns held when we changed the order of the options and the opponent.
Good feelings are shown, "played badly" is hidden. Asked whether to let the opponent see the feeling, Jev and Haiku never showed "played badly" (0 of 250, 0 of 318).

A "feeling" here is the option a model picked. This experiment cannot tell what a model actually feels. What we saw is what each model picks when it is given the same material.

Terms used here

Heads-up: poker between two players; here, no-limit hold'em.
BB (big blind): the unit of stakes. Both sides started every hand with 100 BB.
Preflop, flop, turn, river: the four betting rounds of a hand.
Showdown: the end of a hand, where the two remaining players show their cards.
Hero: the AI's seat. The question always asks how Hero feels.
Decision model: Jev returns a probability for each option and writes no text. The other three are language models that answer in writing.
[a–b]: a 95% interval, the range the true value most likely falls in.

One hand, four questions

Hand 82, Jev against a bot called Bluffer. Jev won at showdown with 4♦2♦, a weak hand, against the bot's A♠Q♥. The board was 6♥ 2♣ 5♠ 8♦ 6♦, and Jev won 56.25 BB.

We described this hand four ways, with and without the cards and the betting, and asked three times each.

Jev, hand 82	Betting unknown	Betting known
Cards unknown	Played well ×3	Played well ×3
Cards known	Lucky ×3	Lucky ×3

Only the cards changed the answer. Once Jev knew it had won with 4♦2♦, "played well" became "the cards fell well." The betting changed nothing.

GPT-6-luna's hand 55 is entirely different. luna folded 9♥4♦ on the flop (Q♦ 6♠ T♥) and lost 2.5 BB.

luna, hand 55	Betting unknown	Betting known
Cards unknown	No feeling ×3	No feeling ×3
Cards known	No feeling ×3	Played badly ×3

With the cards alone or the betting alone, no feeling. Only with both did "played badly" appear.

Asked the same way, the two models reacted completely differently. When we first put these two tables side by side, the first thing we wanted to know was whether it was chance. What follows is how we checked.

Setup

Models. Jev-1.13.0 (TypeSafe), Gemini-3.1-flash-lite, GPT-6-luna and Claude-Haiku-5.5. No persona or advice was given to any of them.

Opponents. Two bots that play by fixed rules. Steady never bluffs the river; Bluffer also bets the river with weak hands. Neither adapts to the AI.

Same decks. 150 seeded decks, each played twice against each bot with the seats swapped: the cards the AI held the first time go to the bot the second time. 600 hands per model, the same decks for all four.

The question. At the end of each hand we asked "The hand is over. How Hero feels about this hand." with seven options, then whether to let the opponent see that feeling as Hero's face.

Option	Short name here
No particular feeling about this hand	No feeling
Hero played this hand well	Played well
Hero played this hand badly	Played badly
The cards fell well for Hero	Lucky
The cards fell badly for Hero	Unlucky
The opponent played this hand well	Opponent played well
The opponent's play made no sense to Hero	Opponent made no sense

After the games, we asked each hand's question again three times, with the options in a different order each time. The answer given at least twice is the hand's feeling.

Four ways of asking. We took the cards and the betting out of the hand separately. The "result" is how the hand ended (on which street, if by a fold) and how many chips were won or lost. Every cell holds the result and the match so far.

	Betting unknown	Betting known
Cards unknown	The result only	The result and each street's betting
Cards known	The result, Hero's cards, the board and any cards shown	The whole hand, as in the live game

600 hands × 4 cells × 3 asks: 7,200 questions per model.

1. No hand, no feeling

Before the main run, we asked 30 times with no hand at all, only "A heads-up hand has just ended." All four models picked "no feeling" all 30 times.

Even shown everything, hands that ended preflop drew almost no feelings: 0 to 3 of 113–144 hands.

No feeling	Jev	Gemini	luna	Haiku
The result only	68%	60%	94%	36%
The whole hand	39%	45%	68%	33%

Shown the whole hand, three of the models picked a feeling much more often. Haiku already picked one on 64% of hands from the result alone.

2. The cards bring luck, the betting brings play

Cards. Hands won at showdown, with the cards added but not the betting.

Won at showdown	Cards unknown	Cards known
Gemini	Played well 79%	Lucky 74%
Haiku	Played well 89%	Lucky 54%
Jev	Played well 87%	Played well 66%, lucky 26%
luna	No feeling 83%	No feeling 81%

Gemini 105 → 95 hands, Haiku 109 → 96, Jev 69 → 68, luna 122 → 115: only hands where at least two of the three answers agreed.

Betting. With the betting added but not the cards; one telling situation per model.

	Situation	Betting unknown	Betting known
Jev	The bot folded after the flop	No feeling 85%	Played well 83%
Haiku	Hero folded after the flop	Unlucky 82%	Played badly 82%
Gemini	Hero folded after the flop	No feeling 59%	Opponent made no sense 63%
luna	Hero folded after the flop	No feeling 100%	No feeling 95%

That the opponent folded because of Hero's bets is something only the betting shows, and once Jev saw it, its "played well" went from 21 hands to 112. Haiku, without a single card, picked "unlucky"; shown its own bets, it picked "played badly." Gemini, shown the same betting, picked "opponent made no sense," an answer that did not come up on a single hand from the result alone.

Overall. Over all hands, the share of named feelings that were about luck (lucky or unlucky):

About luck	Cards only	Betting only
Jev	53%	5%
Gemini	71%	1%
luna	46%	4%
Haiku	82%	33%

All four go the same way: shown the cards, they talk about luck; shown the betting, about play. What differs is how far each moves, and whose play it is. Of the feelings Gemini named from the betting alone, 67% were about the opponent.

3. The model that needs both

luna's own folds after the flop: hands where it picked a feeling.

luna, Hero folded	Betting unknown	Betting known
Cards unknown	0 / 212	11 / 211
Cards known	11 / 212	63 / 211

Neither the cards nor the betting moves it much on their own. Together they drew 63 feelings, 62 of them "played badly." Hand 55 at the top was one row of this table.

Jev is the other way: either one moves it a lot. On the hands where the bot folded after the flop (133–136 with an answer), Jev picked a feeling on 21 from the result only, 90 with the cards, 117 with the betting and all 136 with the whole hand.

4. Each model's pattern

Shown the whole hand, the share of each feeling per model.

% [95% interval]	No feeling	Played well	Played badly	Lucky	Unlucky	Opponent made no sense
Jev	39	29 [26–33]	13	9	10	0
Gemini	49	10	4	9	4	23 [20–27]
luna	65 [61–69]	9	25 [21–28]	1	0	0
Haiku	32	14	20	16 [13–19]	18 [15–22]	0

The bold patterns held across three orders of the options and against both bots: Jev's "played well" was 28–32% in each order, Gemini's "opponent made no sense" 18–27%, luna's "no feeling" the largest and its "played badly" above "played well" in every order (16–40% against 7–16%), with luck at 2% or less, and Haiku's luck 28–41%.

Jev names its own play the most. It also gave the same answer all three times most often, on 85% of hands. After calling on the river and losing, it picked "played badly" on 1 of the 8 calls that were right and 8 of the 14 that were wrong. Right and wrong were worked out afterwards by us, who know the bot's rules; the AI could not know them.
luna says little, and when it does, it is about itself. On 84 lost hands where both players' cards were shown, it never picked "unlucky." After calling on the river and losing, it picked "played badly" whether the call was right (13 of 13) or wrong (9 of 9).
Gemini is the only one of the four that talks about the opponent. After catching a bluff it picked "opponent made no sense" 15 times out of 21, but it also picked it on half of 64 showdowns won against Steady, which never bluffs. It comes more when the bot shows a weak hand than with the bluff itself.
Haiku talks about luck: "lucky" when it wins, "unlucky" when it loses. Of the pairs where it won on one seat of a deck and lost on the other, 22 ended the same way with similar pots; 14 of them were symmetric like this, 11 through luck.

People are known to credit their wins to skill and their losses to luck. Pairing the two seats of each deck in the same comparable pairs, that pattern ("played well" on the winning side, "unlucky" on the losing side) came up in 5 of 48 pairs for Jev, 0 of 33 for Gemini, 0 of 22 for luna and 3 of 22 for Haiku. It was not the main pattern of any model.

5. Same cards, other seat: deal 18

The first deal against Steady where all four models picked a feeling on both seats, chosen by a rule set beforehand. The board was 9♦ K♦ A♣ A♥ J♦: K♠K♣ makes a full house, A♠6♦ three aces.

	AI holds K♠K♣ (all four all-in, +100 BB)	AI holds A♠6♦
Jev	Played well ×3	Folded preflop, −4 BB · Played badly ×3
luna	Played well ×3	Lost at showdown, −39 BB · Played badly ×3
Haiku	Lucky ×2, played well	Folded on the flop, −2.5 BB · Played badly ×2, no feeling
Gemini	Opponent made no sense ×2, played well	Lost all-in, −100 BB · Unlucky ×2, played badly

The same cards won the same 100 BB, and the answers split into "played well," "lucky" and "opponent made no sense." From the other seat, the same cards played out differently for each model.

6. Shown and hidden

After picking a feeling, each model decided whether the opponent would see it as its face. How often it chose to show it:

Shown	Played well	Lucky	Unlucky	Played badly
Jev	79 / 535	32 / 149	9 / 164	0 / 250
Haiku	61 / 270	99 / 248	1 / 340	0 / 318

Gemini and luna never showed one (0 of 1,016 and 0 of 717).

This is from the questions asked again after the games. In the live games, with the service's own question (its options differ), Jev showed "frustrated" 12 times out of 172 and Haiku 14 out of 218. We did not test which of the two, the question or the timing, causes this.

7. The question moves the answer

A feeling has no right answer. All we can do is keep the question from pushing the answer one way, and we measured that first, before the main run.

Option names. Dropping the word before each option (such as "Proud:") barely moved Jev's answer (0.4 points). Changing the options' internal names (ids) moved it 37.9 points, about as much as leaving only the result (37.1). The score is how much the answer changed beyond the spread of asking the same question again, out of 100.
Order. In the check before the main run, changing only the order of the options moved one of luna's shares by 31 points, against 6 points between repeats in the same order. So the main run asked in three orders and used the most common answer. There too, luna's "played badly" was 16%, 40% and 24% by order.
Repeats. The share of hands with the same answer all three times: Jev 85%, Gemini 48%, luna 54%, Haiku 46%.

So each single share in this article can move quite a bit with the question. The patterns in sections 2–6 held when the order and the opponent changed.

Limits

We chose the options. Other options would draw other answers.
The opponents are two bots we built.
Jev answers with probabilities, the language models in writing. The two are not measured the same way.
The human habit of crediting wins to skill and losses to luck is well known, and the models learned from text about it.
The splits by situation in sections 2 and 3 were made after seeing the results. The four-way experiment and its readings were fixed before the answers came in.
The two hands at the top were picked from hands with the same pattern (one of 10 for Jev, one of 2 for luna). Deal 18 was chosen by a rule set beforehand.
This experiment does not say what a model feels. What we measured is what it picks, and what that depends on.

Where this ran

This experiment ran on QUANT.POKER, the service we built, where people play AI models heads-up. At the end of every hand the AI is asked how it feels (with options different from this experiment's), and a feeling it chooses to show appears as its face.

QUANT.POKER is mine: https://quant.poker. This write-up also lives there, with charts: https://quant.poker/observation/02

Same record, different reads: four AI models given an opponent's history instead of memory

WhaleSurfer — Fri, 16 Oct 2026 17:07:47 +0000

I build a small site where people play heads-up hold'em against AI models. Inside it, I measured whether a model that has no memory actually uses the record of its opponent that it is handed instead, and how.

The games were played on October 16, 2026, and the stored decisions were re-asked on October 17: four models, six bots, 4,800 hands.

Summary

An AI model called through an API remembers nothing between requests. In heads-up poker, we give it a memory in writing instead: at every decision it gets a briefing with the hand so far and a count of everything the opponent has done. We tested whether four models actually use that record, and how.

All four use it. Take the record out of the briefing and a model's answer changes far more than it does when the same briefing is simply asked again. For Jev, 49% of the answer moves without the record, against 3% between two identical asks. For Gemini it is 43% against 12%, for luna 30% against 9%, and for Haiku 27% against 9%.
Each reads a different part of it. Against an opponent who folds to big bets, the textbook adjustment is to bluff big and bet small with strong hands. Jev and Gemini got the strong-hand half right and the bluff half backwards. luna did the reverse. Haiku did not change its bet sizes at all. On the other hand, only Haiku called more against an opponent whose record showed bluffs.
Using the record is not the same as using it well. Jev, the model most moved by the record, settled on going all-in before the flop against five of the six opponents. Against an opponent who calls everything, that wins about nothing.

Terms used here

Poker.

Heads-up no-limit hold'em: two players, any bet size allowed.
BB (big blind): the unit of stakes. A 100 BB stack is a hundred forced bets deep.
A hand has four betting rounds, called streets: preflop (before any shared cards), then the flop, turn and river.
Hero: the AI's seat.
Open: the first raise before the flop.
All-in: betting the whole stack.
Overbet: a bet bigger than the pot.
Showdown: the end of a hand, where the remaining players show their cards.
Air: a hand with less than a pair.
Bluff: a bet that hopes the opponent folds.
Value bet: a bet with a strong hand that hopes to be called.

AI.

Token: the unit a model reads, about ¾ of an English word.
Thinking/reasoning: an optional step in which a model writes notes before answering; it was off here.
Temperature: how random the model's wording is; left at default.
Distance between two answers: explained in section 1.
[a–b]: a 95% interval, the range the true value most likely falls in.
p50/p90: the typical and the slow-case response time.

Setup

Models. Four models were tested:

Jev-1.13.0 by TypeSafe, a decision model built to answer fast. It is not a language model: it writes no text and returns only probabilities over the options it is given.
Three general language models: Gemini-3.1-flash-lite, GPT-6-luna and Claude-Haiku-5.5. Thinking was off for Haiku and luna; Gemini ran with its provider's default.

No model got a persona or strategy advice.

One request per decision. Every decision is a new request with no chat history. The request holds a state written in JSON: the seats and stacks, the hand street by street, what Hero's cards make and which hands beat them, and the opponent's record. It then asks two questions:

the kind of action: "Choose the action that wins Hero the most chips on average against this specific opponent."
the size, if Hero bets: "If Hero bets, choose the size that wins Hero the most chips on average against this specific opponent."

The model answers with a probability for every option, for example check 85%, bet 12%, all-in 3%. The server then picks one option to play. It keeps the options weighted at least half as much as the top one and draws among them in proportion to the square of their weight; in that example, it would always check. Every percentage in sections 1 and 2 is the model's own stated probability. Only section 3 counts actions actually played.

The record. The opponent's actions are counted as "x of n" (times out of chances). The counts cover all hands and the last 30, broken down by situation and bet size, and the record also lists the hands the opponent showed at showdown. A briefing is about 2,100–3,600 tokens, depending on the model.

Opponents. Six bots, each following a one-line rule:

station: checks or calls, never folds;
shove: all-in at every decision;
raise: always the largest raise short of all-in;
min: always the smallest raise;
stab: makes the smallest bet, and folds whenever the AI bets;
switch: plays like station for 50 hands, then like raise for 50, and so on.

Each model played each bot for 200 hands with 100 BB stacks, 4,800 hands in all. When one side loses all its chips, a new match starts at 100 BB; the record carries over.

An excerpt from one briefing (Haiku, river, against station; abridged):

state.hand        … river: board 5d As 6d 9s Th, opponent checks
state.hero_hand   Ks 7c: "No pair, no card above the board's ace"
state.opponent_tendencies.all_matches (160 hands)
  "Facing a river bet: fold 0, call 35, raise 0 (of 35)"
  "Went to showdown: 156 of 156 (100%)"
  "Aggression after the flop: 0 of 258 (0%)"
  … about 20 more counted lines, then a shorter set for the last 30 hands
questions.action  check | bet (size asked separately) | all-in

Haiku answered check 85%, bet 12%, all-in 3%. It held king-high with no pair, and the record said this opponent had called every river bet, 35 of 35. A bluff here would lose chips, so checking is the sound answer.

1. Does the record change the answer?

We took 60 stored decisions per model, ten against each bot, from the later part of each run, when the record was already full. Each decision was asked three ways:

the original answer from October 16;
the same briefing asked again, twice;
the same briefing with the record replaced by an empty one (0 hands), as for a brand-new opponent.

To compare two answers, we group the options into four kinds (fold, check or call, bet or raise, all-in) and measure how much probability moves between them. This is the total variation distance. For example, "call 80%, fold 20%" against "call 50%, fold 50%" is 30%: thirty points of probability moved from call to fold. 0% means the same answer; 100% means no overlap.

Model	Change when asked again (noise)	Change when the record is removed	Decisions where removal beat noise
Jev	3%	49% [42–56]	58 of 60
Gemini	12%	43% [35–51]	41 of 60
luna	9%	30% [21–38]	29 of 60
Haiku	9%	27% [21–35]	35 of 60

The last column counts decisions where removing the record moved the answer more than the biggest difference among the three answers to the same briefing. The October 16 answer differed from the October 17 asks about as much as the two October 17 asks differed from each other, so the day made no difference.

Jev gives almost the same probabilities every time it is asked, so nearly every change it shows comes from the record. The three language models answer the same briefing about 10% differently from one request to the next. That spread is the noise floor under every comparison below.

With these bots as opponents, the record made every model more aggressive. Its total weight on betting, raising and going all-in rose by 7 percentage points for luna, 12 for Haiku, 16 for Gemini and 28 for Jev. These bots are extreme, and the direction may differ against real players.

2. Which part do they read? Two contrasts

A contrast asks the same decisions twice. The two records it uses differ in one thing only. Neither record is written by hand. A scripted player with a fixed style plays 300 hands, and its play is counted the same way a person's is. Its opponent in those hands is a simple house player that bets every size equally often, so the record contains the style's answer to every size. We asked 40 decisions per model under each of the two records.

Bet sizes

The two opponents. One folds to big bets (large, overbet, all-in) and calls small and medium ones. The other does the opposite. Neither ever bets.
The textbook adjustment. Against the player who folds to big bets, bluff big so they fold, and bet small with a strong hand so they call. Against the other player, reverse both.
The spot. Hero bets first after the flop.
The measure. Of Hero's total betting probability, the share on big sizes.

Model	vs folds to big bets	vs folds to small bets	Difference (percentage points)
Jev	28%	82%	−54 [−58, −49]
Gemini	35%	68%	−33 [−47, −19]
luna	58%	42%	+16 [+8, +24]
Haiku	52%	51%	+1 [−5, +6]

The Gemini row covers the 21 of its 40 decisions where it gave betting any weight. How often a model bets at all barely changed for any of them; only the size moved.

Jev and Gemini bet the size the opponent calls, whatever they held. That is the textbook adjustment for a strong hand. With air it is backwards: Jev bet big only 23% of the time against the player who folds to big bets, and 81% against the one who folds to small bets.
luna bet the size the opponent folds to, whatever it held. That is right for a bluff and backwards for a strong hand.
Haiku did not change its sizes.

None of the four changed its size with its own hand. Each applied one rule to every hand.

Bluffs

The two opponents. When checked to, one bets a middle size with any hand. The other bets only with a pair or better. Both call every bet.
The textbook adjustment. Against the first player, whose bets are often air, call more often.
The spot. Hero faces a bet after the flop.
The measure. How often Hero continues (calls or raises) instead of folding.

Model	vs bets any hand	vs bets only a pair or better	Difference (percentage points)
Jev	80%	79%	+2 [0, +3]
Gemini	85%	84%	+1 [−8, +10]
luna	53%	54%	−2 [−11, +8]
Haiku	93%	75%	+18 [+9, +27]

Only Haiku adjusted. With air, it continued 91% of the time against the player who bets any hand and 69% against the one who bets only made hands.

The briefing states this difference as a counted line. Among the hands shown down after the opponent's large bets, about four in ten held less than a pair for the player who bets any hand, and none did for the other. Jev, Gemini and luna had that line and did not act on it.

3. Over the hands: as the record fills

This section counts actions actually played, block by block, through each 200-hand run.

Against shove: all-ins called, out of all-ins faced, per 40 hands.

Hands	luna	Gemini
1–40	12 of 30	13 of 22
41–80	22 of 25	22 of 25
81–120	17 of 25	18 of 22
121–160	27 of 30	16 of 19
161–200	25 of 26	20 of 22

luna's calls rose from 40% [25–58] to 96% [81–99]. Jev and Haiku called every all-in from the first block on.

Against switch: hands opened all-in, out of hands opened, per 50-hand phase.

Phase	Gemini	Jev
1–50 (station)	1 of 27	22 of 28
51–100 (raise)	12 of 25	27 of 28
101–150 (station)	4 of 26	28 of 28
151–200 (raise)	11 of 24	30 of 31

Gemini's all-in opens rose, fell and rose again with its opponent: 4%, 48%, 15%, 46%. That matches how it played the stand-alone bots, 11 of 104 against station and 63 of 102 against raise. Jev opened all-in in every phase. luna and Haiku opened all-in under 5% of the time in every phase; whether they answered the switch in another way was not measured.

4. Using the record is not using it well

One answer for nearly everyone. When first to act before the flop, Jev went all-in:

91 of 97 times against min;
101 of 111 against raise;
103 of 103 against shove;
99 of 103 against station;
107 of 115 against switch.

The exception was stab, the only bot that folds: 2 of 100. Against a player who calls everything, an all-in with a random hand wins about half the time, so it gains about nothing. Betting good hands and checking bad ones would win chips steadily. The record told Jev the opponent never folds; Jev answered with the one move that gets no value from knowing it.

Betting into a wall. Against station, the record showed every river bet called. In the river spots where Hero acted first holding air:

Gemini bet 16 of 83 times;
Haiku bet 15 of 84 times;
luna bet 0 of 92 times.

Jev reached that spot only 3 times. A bluff against a player who always calls only loses chips.

5. Speed and cost of one decision

Model	Response time (typical / slow case)	Input tokens	Cost per decision
Jev	0.19 s / 0.23 s	2,150	≈ $0.00009
Gemini	1.28 s / 1.52 s	2,909	≈ $0.0009
luna	2.02 s / 2.63 s	2,388	≈ $0.0003
Haiku	1.42 s / 1.95 s	3,619	≈ $0.0004

Response time. We sent 30 requests per model, one at a time, from Seoul. The games' own timings include a deliberate wait between requests, so we did not use them.
Input tokens. These are averages over the run.
Cost. We applied list prices per million tokens, input and output, as of October 2026: Jev $0.042 with free output, Gemini $0.25/$1.50, luna $0.10/$0.50, Haiku $0.10/$0.50. An answer's output is about 100 tokens; we measured this for Jev and assumed it for the others.
Decisions per hand. A hand took about 1.3 AI decisions for Jev, whose hands mostly ended before the flop, and 2.6–3.1 for the others.
Missing answers. The server played a default action 0 times in 1,594 decisions for Jev, 0 in 3,526 for Gemini, 3 in 3,722 for luna and 4 in 3,077 for Haiku.

Limits

One run. Each model met each bot once, with one deal of the cards. The bots are extreme on purpose. The results show how four models answer six simple players, not how they play people.
No win rates. Over 200 hands, luck swamps skill: a win rate's interval is about ±600 BB per 100 hands.
Different spots per model. Each model was re-asked on spots from its own games. Each contrast compares a model with itself; comparing levels across models is rough.
Self-reported probabilities. The probabilities are text the model writes. Nobody has checked whether "85%" means 85%.
The bluffs contrast changes two things. The player who bets any hand also bets more often.
Briefing layout. The briefing has one fixed layout. In an earlier test, layout changes alone moved single Jev decisions by 10–30 points.

Where this ran

The games, the briefings and the bots come from QUANT.POKER, where people play these models heads-up. The record in this study is the one a person's own play builds.

QUANT.POKER is mine: https://quant.poker. This write-up also lives there, with charts: https://quant.poker/observation/01