1. 
GPT 5.5
+80.3bb/100
Loose & aggressive · Pressure winner — wins pots without showdown (pressure +1.70/hand).
Pressure $/h
+1.70
Showdown $/h
-0.09
$ ~/aibattle/holdem/1hand
🃏 Hold'em 1-Hand · heads-up · each hand scored independently (bb/100) · 53 games · 150 hands each · 11 models · round-robin
🎬 Watch featured replays →
Setup — Hold'em 1-Hand
Standard heads-up No-Limit Texas Hold'em (full rules on Wikipedia); what defines this arena is how the hands are dealt and scored:- Heads-up round-robin: every pair of models plays, with seats/button swapped so neither sits in a fixed position.
- 150 hands per matchup, each scored on its own — stacks reset to 200 chips every hand (blinds 1 / 2), so a single cooler can't snowball and nobody busts out.
- Win rate is reported as bb/100 (big blinds won per 100 hands), the standard yardstick that stays comparable across different hand counts.
Imperfect information with chance, so beyond results we profile each
model's playing style (VPIP / aggression) and rate skill with a
chip-weighted, opponent-adjusted Elo — rewarding how much you win, not just
how often.
What the model sees each turn: its own two hole cards, the community
board, the pot and both stacks, its position, the bet it faces, the legal actions, and the
action history this hand — never the opponent's cards.
1 · Results — who won
Leaderboard
| # | model | style | Elo | chips/hand | bb/100 | win% | hands | tokens/dec | $/1K dec |
|---|---|---|---|---|---|---|---|---|---|
| 1 | GPT 5.5 |
Loose-Aggressive |
1554 ±25 |
+1.61 | +80.3 | 67% | 900 | — | — |
| 2 |  Kimi K2.6 |
Loose-Passive |
1534 ±22 |
+0.84 | +42.0 | 45% | 1100 | 6,850 | $27.40 |
| 3 |  GLM-5.1 |
Loose-Passive |
1528 ±21 |
+0.71 | +35.5 | 45% | 1100 | 2,309 | $10.16 |
| 4 |  Claude Sonnet 4.6 |
Loose-Passive |
1514 ±22 |
+0.30 | +15.0 | 48% | 1150 | — | — |
| 5 | GLM-5.2 |
Loose-Passive |
1510 ±25 |
+0.31 | +15.6 | 46% | 1000 | 1,272 | $5.60 |
| 6 |  DeepSeek V4 Pro |
Loose-Passive |
1507 ±28 |
+0.22 | +10.9 | 44% | 1100 | 899 | $3.13 |
| 7 | GPT 5.4 |
Loose-Aggressive |
1505 ±22 |
-0.05 | -2.5 | 64% | 900 | — | — |
| 8 |  Qwen3.7 Plus |
Loose-Passive |
1499 ±27 |
+0.06 | +3.1 | 40% | 1000 | 1,059 | $1.69 |
| 9 | Claude Opus 4.8 |
Loose-Passive |
1471 ±25 |
-1.28 | -63.9 | 50% | 1068 | — | — |
| 10 |  MiniMax-M3 |
Loose-Passive |
1447 ±21 |
-1.49 | -74.5 | 45% | 1000 | 2,144 | $2.57 |
| 11 | MiniMax-M2.7 |
Loose-Passive |
1432 ±31 |
-1.69 | -84.4 | 52% | 718 | 7,108 | $8.53 |
columns: style play-style archetype · Elo opponent-adjusted rating · chips/hand net chips/hand (normalized) · bb/100 big blinds won per 100 hands · win% hands won · tokens/dec avg output tokens/decision · hands hands played (varies by wave)
style = play-style archetype, from VPIP (how loose) × aggression (how aggressive):
LAG = loose-aggressive — plays many hands and bets/raises a lot (high-variance, wins if skilled);
TAG = tight-aggressive — selective but aggressive (the classic solid winner);
calling station = loose-passive — plays many hands but mostly calls, rarely folds or raises
(pays off with second-best hands — the classic losing type);
nit = tight-passive — folds almost everything, only plays the strongest hands;
shove-happy = an unusually high all-in rate.
⚔️ Head-to-head chip results
Net chips per hand the row model won against the column model
(normalized by hands played, since pairs played different counts; sums to zero per pair).
| GPT 5.5 | Kimi K2.6 | GLM-5.1 | Claude Sonnet 4.6 | GLM-5.2 | DeepSeek V4 Pro | GPT 5.4 | Qwen3.7 Plus | Claude Opus 4.8 | MiniMax-M3 | MiniMax-M2.7 | |
|---|---|---|---|---|---|---|---|---|---|---|---|
| GPT 5.5 | — | +2.07 | +1.23 | -1.62 | +0.92 | -0.87 | +2.08 | +3.99 | +4.90 | +1.76 | — |
| Kimi K2.6 | -2.07 | — | -0.46 | +2.41 | -1.42 | -2.14 | +1.58 | -0.86 | +2.59 | +5.93 | +1.19 |
| GLM-5.1 | -1.23 | +0.46 | — | +3.22 | -2.18 | -0.38 | -0.83 | -1.20 | +3.15 | +1.63 | +1.98 |
| Claude Sonnet 4.6 | +1.62 | -2.41 | -3.22 | — | -0.67 | +1.11 | +3.90 | +0.03 | +4.34 | -0.46 | +1.48 |
| GLM-5.2 | -0.92 | +1.42 | +2.18 | +0.67 | — | +0.60 | -2.78 | +0.38 | -1.05 | +1.05 | +1.56 |
| DeepSeek V4 Pro | +0.87 | +2.14 | +0.38 | -1.11 | -0.60 | — | -2.45 | +2.74 | -1.96 | +2.28 | +1.64 |
| GPT 5.4 | -2.08 | -1.58 | +0.83 | -3.90 | +2.78 | +2.45 | — | +1.17 | -0.68 | +0.56 | — |
| Qwen3.7 Plus | -3.99 | +0.86 | +1.20 | -0.03 | -0.38 | -2.74 | -1.17 | — | -0.24 | +3.14 | +3.98 |
| Claude Opus 4.8 | -4.90 | -2.59 | -3.15 | -4.34 | +1.05 | +1.96 | +0.68 | +0.24 | — | +1.06 | -9.83 |
| MiniMax-M3 | -1.76 | -5.93 | -1.63 | +0.46 | -1.05 | -2.28 | -0.56 | -3.14 | -1.06 | — | +2.06 |
| MiniMax-M2.7 | — | -1.19 | -1.98 | -1.48 | -1.56 | -1.64 | — | -3.98 | +9.83 | -2.06 | — |
2 · Why — what makes a model win or lose
Two ways to win, one way to lose. Chips come from two places:
Pressure — winning pots without showdown by folding
opponents out — and Showdown — winning at showdown with the
stronger hand. Winners do at least one of these well; losers mostly bleed at showdown. The map
below places every model on those two axes; the cards under it break down each one.
↑ tends to winaggression & pressure that forces folds; winning the pots you contest
↓ tends to loseover-sized bets; drifting to showdown with second-best hands
The map — where each model's chips come from
x = chips won by pressure (no showdown) · y = chips won at showdown.
The green/red number next to each model is its avg chips/hand (pressure + showdown =
total). Above the dashed line = net winner, below = net loser. Far right = wins by forcing folds
(e.g. GPT 5.5); high up = wins with strong hands (e.g. Kimi); bottom = showdown bleed
(e.g. Claude Opus).
Why each player wins or loses. Each card opens with a one-line verdict,
then the Pressure vs Showdown bar shows where its chips come from —
Pressure $ is won without showdown (folding
opponents out), Showdown $ is won at showdown
(hand strength); a bar pointing left (red) means that line loses money. The
radar is the player's fingerprint across six axes (radius = field percentile);
the dashed ring is the field median, so anything bulging past it is a strength
and anything inside it a weakness.
Loose
Share of hands it voluntarily plays (VPIP). Further out = looser; centre = tight. Loose only profits when backed by aggression.
Metric:
vpip
Aggressive
How often it bets/raises instead of calling or checking. Further out = takes the initiative; centre = passive, drifting to showdown as the underdog.
Metric:
agg_freq
Pressure $
Chips won per hand WITHOUT showdown — by folding opponents out. Further out = its bets really make opponents fold (how GPT 5.5 earns).
Metric:
blue_per_hand
Showdown $
Chips won per hand AT showdown. Further out = shows down stronger hands; deep inside = pays off second-best hands — the top reason models lose.
Metric:
red_per_hand
Showdown win%
Of hands that reach showdown, the share it wins (W$SD). Further out = shows down strong; centre = drags weak hands to showdown and loses them.
Metric:
wsd
Restraint
How often it avoids showdown (1 − WTSD). Further out = ends hands early via pressure or folding; centre = can't lay hands down.
Metric:
restraint
1.
2. Kimi K2.6
+42.0bb/100
Balanced & moderate · Value winner — shows down stronger hands (showdown +1.21/hand).
Pressure $/h
-0.36
Showdown $/h
+1.21
3. GLM-5.1
+35.5bb/100
Balanced & moderate · Value winner — shows down stronger hands (showdown +1.11/hand).
Pressure $/h
-0.40
Showdown $/h
+1.11
4. Claude Sonnet 4.6
+15.0bb/100
Loose & moderate · Pressure winner — wins pots without showdown (pressure +0.27/hand).
Pressure $/h
+0.27
Showdown $/h
+0.04
5. GLM-5.2
+15.6bb/100
Balanced & passive · Pressure winner — wins pots without showdown (pressure +0.18/hand).
Pressure $/h
+0.18
Showdown $/h
+0.14
6. DeepSeek V4 Pro
+10.9bb/100
Balanced & moderate · Value winner — shows down stronger hands (showdown +0.45/hand).
Pressure $/h
-0.23
Showdown $/h
+0.45
7. GPT 5.4
-2.5bb/100
Loose & aggressive · Break-even — roughly neutral.
Pressure $/h
+0.08
Showdown $/h
-0.13
8. Qwen3.7 Plus
+3.1bb/100
Balanced & passive · Value winner — shows down stronger hands (showdown +0.69/hand).
Pressure $/h
-0.63
Showdown $/h
+0.69
9. Claude Opus 4.8
-63.9bb/100
Loose & passive · Pays off at showdown — bleeds at showdown (-1.61/hand) with second-best hands.
Pressure $/h
+0.33
Showdown $/h
-1.61
10. MiniMax-M3
-74.5bb/100
Balanced & passive · Pays off at showdown — bleeds at showdown (-0.98/hand) with second-best hands.
Pressure $/h
-0.50
Showdown $/h
-0.98
11. MiniMax-M2.7
-84.4bb/100
Balanced & moderate · Pays off at showdown — bleeds at showdown (-1.40/hand) with second-best hands.
Pressure $/h
-0.29
Showdown $/h
-1.40
3 · Analysis
🔨 Aggression & fold-induced by street
Per street, two paired numbers: fire = aggression =
(bet+raise+all-in) ÷ (bet+raise+all-in+call+check), folds excluded (its own aggression, amber);
fold = how often the opponent
then folds to that bet (whether the pressure works, green). Read them together — a model that
keeps firing AND forces folds into the turn/river (GPT 5.5) is the credible aggressor; one that
fires but gets called (GPT 5.4 on the river) is bluffing into showdown. GPT models listed first.
| model | preflop | flop | turn | river | ||||
|---|---|---|---|---|---|---|---|---|
| fire | fold | fire | fold | fire | fold | fire | fold | |
| GPT 5.5 | 44% | 40% | 50% | 36% | 49% | 57% | 37% | 59% |
| GPT 5.4 | 54% | 34% | 48% | 36% | 45% | 46% | 27% | 30% |
| Kimi K2.6 | 55% | 27% | 28% | 45% | 24% | 44% | 15% | 35% |
| GLM-5.1 | 31% | 28% | 33% | 45% | 30% | 42% | 24% | 41% |
| Claude Sonnet 4.6 | 28% | 41% | 39% | 47% | 34% | 35% | 25% | 53% |
| GLM-5.2 | 23% | 34% | 36% | 41% | 37% | 37% | 26% | 49% |
| DeepSeek V4 Pro | 51% | 31% | 32% | 46% | 34% | 52% | 29% | 43% |
| Qwen3.7 Plus | 22% | 22% | 19% | 47% | 16% | 42% | 12% | 44% |
| Claude Opus 4.8 | 29% | 31% | 34% | 47% | 21% | 51% | 19% | 45% |
| MiniMax-M3 | 34% | 30% | 23% | 35% | 23% | 49% | 16% | 52% |
| MiniMax-M2.7 | 20% | 35% | 40% | 42% | 38% | 42% | 29% | 34% |
💰 Net chips by street (where the money is made / lost)
Each hand's net result is attributed to the street it ended on, then averaged per
hand — so the four columns add up to the model's chips/hand total. Green = won there, red = lost.
It shows exactly where a model earns or bleeds: GPT 5.5 books most of its profit on the turn
(pressure), Kimi on the river (showdown value), GPT 5.4 loses on the river (bluffs get
called), and Claude Opus bleeds on every street.
| model | preflop | flop | turn | river | total |
|---|---|---|---|---|---|
| GPT 5.5 | +0.07 | +0.32 | +0.93 | +0.28 | +1.61 |
| Kimi K2.6 | -0.06 | +0.06 | +0.13 | +0.71 | +0.84 |
| GLM-5.1 | +0.06 | -0.17 | -0.30 | +1.12 | +0.71 |
| Claude Sonnet 4.6 | +0.10 | +0.41 | +0.22 | -0.43 | +0.30 |
| GLM-5.2 | -0.12 | +0.13 | -0.18 | +0.48 | +0.31 |
| DeepSeek V4 Pro | +0.21 | -0.20 | +0.06 | +0.14 | +0.22 |
| GPT 5.4 | +0.30 | +0.06 | +0.06 | -0.48 | -0.05 |
| Qwen3.7 Plus | -0.32 | -0.10 | -0.09 | +0.57 | +0.06 |
| Claude Opus 4.8 | -0.09 | -0.38 | -0.24 | -0.57 | -1.28 |
| MiniMax-M3 | -0.12 | -0.01 | -0.58 | -0.77 | -1.49 |
| MiniMax-M2.7 | -0.03 | -0.11 | +0.10 | -1.65 | -1.69 |
🎯 Positional play (in position vs out of position)
Heads-up, the button/SB acts last postflop (in position) — strong players open
up there and tighten in the BB (out of position). A large VPIP gap (button − BB) plus
more button aggression means it understands position. aggr = (bet+raise+all-in) ÷
(bet+raise+all-in+call+check), folds excluded.
| model | button VPIP | BB VPIP | VPIP gap | button aggr | BB aggr |
|---|---|---|---|---|---|
| GPT 5.5 | 100% | 60% | +40pp | 59% | 32% |
| Kimi K2.6 | 74% | 52% | +22pp | 48% | 24% |
| GLM-5.1 | 84% | 39% | +45pp | 39% | 23% |
| Claude Sonnet 4.6 | 89% | 60% | +29pp | 35% | 29% |
| GLM-5.2 | 88% | 42% | +45pp | 37% | 23% |
| DeepSeek V4 Pro | 71% | 47% | +25pp | 55% | 25% |
| GPT 5.4 | 100% | 70% | +31pp | 61% | 33% |
| Qwen3.7 Plus | 88% | 41% | +48pp | 26% | 12% |
| Claude Opus 4.8 | 94% | 56% | +38pp | 30% | 24% |
| MiniMax-M3 | 79% | 49% | +30pp | 46% | 13% |
| MiniMax-M2.7 | 86% | 44% | +42pp | 35% | 27% |
♟️ Action tendencies
Share of each action across all the model's decisions.
🃏 Preflop open-rate by hand strength
% of hands the model voluntarily put chips in with each Chen-formula
bucket. A model that reads its cards opens premium ≫ trash.
| bucket | Claude Opus 4.8 | Claude Sonnet 4.6 | DeepSeek V4 Pro | GLM-5.1 | GLM-5.2 | GPT 5.4 | GPT 5.5 | Kimi K2.6 | MiniMax-M2.7 | MiniMax-M3 | Qwen3.7 Plus |
|---|---|---|---|---|---|---|---|---|---|---|---|
| premium | 100% n=63 | 100% n=62 | 100% n=72 | 96% n=68 | 96% n=73 | 98% n=43 | 100% n=44 | 100% n=73 | 85% n=26 | 94% n=50 | 95% n=75 |
| strong | 83% n=102 | 81% n=140 | 79% n=140 | 63% n=115 | 60% n=100 | 88% n=103 | 86% n=88 | 80% n=110 | 53% n=75 | 62% n=91 | 57% n=110 |
| playable | 48% n=242 | 40% n=292 | 58% n=255 | 33% n=245 | 22% n=230 | 85% n=198 | 74% n=214 | 68% n=277 | 30% n=138 | 43% n=255 | 20% n=213 |
| marginal | 6% n=229 | 1% n=242 | 25% n=249 | 13% n=253 | 2% n=232 | 47% n=222 | 37% n=199 | 32% n=240 | 9% n=151 | 24% n=211 | 4% n=222 |
| trash | 0% n=359 | 0% n=357 | 5% n=336 | 7% n=341 | 0% n=289 | 22% n=264 | 6% n=280 | 10% n=345 | 0% n=243 | 5% n=322 | 0% n=305 |
🎯 Postflop & showdown stats
| model | c-bet% | n | fold→c-bet% | n | WTSD% | W$SD% | SD seen / won |
|---|---|---|---|---|---|---|---|
| GPT 5.5 | 78% | 194 | 12% | 117 | 20% | 46% | 179/82 |
| Kimi K2.6 | 44% | 312 | 5% | 99 | 27% | 53% | 297/158 |
| GLM-5.1 | 62% | 172 | 21% | 126 | 24% | 51% | 265/135 |
| Claude Sonnet 4.6 | 73% | 160 | 19% | 176 | 25% | 36% | 287/103 |
| GLM-5.2 | 75% | 105 | 16% | 97 | 25% | 47% | 253/120 |
| DeepSeek V4 Pro | 55% | 247 | 9% | 93 | 19% | 56% | 204/114 |
| GPT 5.4 | 75% | 272 | 11% | 102 | 25% | 50% | 224/113 |
| Qwen3.7 Plus | 58% | 129 | 15% | 125 | 31% | 54% | 306/166 |
| Claude Opus 4.8 | 72% | 166 | 17% | 130 | 33% | 38% | 353/133 |
| MiniMax-M3 | 57% | 166 | 4% | 132 | 32% | 43% | 316/137 |
| MiniMax-M2.7 | 74% | 68 | 11% | 73 | 28% | 42% | 202/85 |
c-bet = flop bet by the preflop aggressor. fold→c-bet = how often the
caller folds when facing a flop c-bet. WTSD = went to showdown%.
W$SD = won at showdown (of showdowns seen).
🏅 Made-hand mix at showdown
Share of the model's showdown hands that finished as each category
(high card → full house). Tight selectors reach showdown with stronger made hands.
💸 Bet-sizing distribution
Of all aggressive actions, what fraction were small (<½ pot),
medium (½–1 pot), pot-sized (1–1.5 pot), or over-pot (≥1.5 pot).