1. 
GPT 5.5
65%match win
Strong — aggressive; high-variance (busts out a lot); fights back when behind.
bust-win 10%cap-win 55%cap-lose 18%bust-out 17%
$ ~/aibattle/holdem/match
🃏 Hold'em Match · Heads-up · 20-60 matches/pair · 62 pair logs · 3020 total matches · up to 30 hands/match · stacks carried, match-level winner · primary metric: match win rate
🎬 Watch featured replays →
Setup — Hold'em Match
Standard heads-up No-Limit Texas Hold'em (full rules on Wikipedia); the difference from 1-Hand is that here a whole sit-and-go match is the unit, not a single hand:- Heads-up sit-and-go: both start with 200 chips (blinds 1 / 2) and play until one is busted, or until a cap of up to 30 hands.
- Stacks carry across hands within a match — winning chips early creates a real lead, so position and stack pressure matter.
- 20-60 matches/pair, seats swapped where available; the match winner is whoever busts the other (or leads at the cap).
Win-or-lose by design — chips don't count past the match outcome —
so the Elo rates match wins/losses, opponent-adjusted. Match win rate is the
headline metric.
What the model sees each turn: the match score (which hand of the cap,
and each side's chips), 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 — never the
opponent's cards.
1 · Results — who won
Leaderboard (ranked by Elo; raw metrics kept for reference)
| # | model | Elo | win% | wins/matches | draws | bust-out% | hands/match | avg win margin | matches | tokens/dec | $/1K dec |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | GPT 5.5 | 1607 ±19 | 65% | 235/360 | 0 | 17% | 26.9 | 138.2 | 360 | — | — |
| 2 | GPT 5.4 | 1579 ±18 | 61% | 220/360 | 0 | 12% | 28.0 | 124.4 | 360 | — | — |
| 3 |  Kimi K2.6 | 1533 ±14 | 56% | 302/540 | 0 | 10% | 27.5 | 151.2 | 540 | 7,166 | $28.66 |
| 4 |  GLM-5.1 | 1530 ±14 | 56% | 300/540 | 0 | 8% | 28.6 | 117.4 | 540 | 3,056 | $13.45 |
| 5 |  DeepSeek V4 Pro | 1515 ±14 | 53% | 287/540 | 0 | 4% | 29.1 | 111.4 | 540 | 1,159 | $4.03 |
| 6 | GLM-5.2 | 1503 ±13 | 51% | 266/520 | 0 | 8% | 28.6 | 148.4 | 520 | 1,454 | $6.40 |
| 7 |  Claude Sonnet 4.6 | 1492 ±16 | 48% | 197/410 | 0 | 15% | 26.6 | 186.6 | 410 | — | — |
| 8 |  MiniMax-M3 | 1443 ±15 | 42% | 217/520 | 1 | 10% | 27.7 | 197.9 | 520 | 2,236 | $2.68 |
| 9 | Claude Opus 4.8 | 1441 ±16 | 40% | 165/410 | 1 | 12% | 26.5 | 206.4 | 410 | — | — |
| 10 | MiniMax-M2.7 | 1432 ±16 | 42% | 167/400 | 2 | 12% | 27.9 | 167.2 | 400 | 6,027 | $7.23 |
| 11 |  Qwen3.7 Plus | 1424 ±15 | 39% | 201/520 | 2 | 5% | 29.0 | 141.3 | 520 | 1,166 | $1.87 |
columns: Elo opponent-adjusted rating · win% matches won · wins/matches won / played · draws matches drawn · bust-out% matches losing all chips · hands/match avg hands/match · avg win margin avg chip margin in wins · matches matches played (varies by wave)
Head-to-head (row's match win % vs column — green = winning, red = losing; raw record below)
| GPT 5.5 | GPT 5.4 | Kimi K2.6 | GLM-5.1 | DeepSeek V4 Pro | GLM-5.2 | Claude Sonnet 4.6 | MiniMax-M3 | Claude Opus 4.8 | MiniMax-M2.7 | Qwen3.7 Plus | |
|---|---|---|---|---|---|---|---|---|---|---|---|
| GPT 5.5 | — | 55%22/40 | 52%21/40 | 62%25/40 | 78%31/40 | 55%22/40 | 68%27/40 | 72%29/40 | 75%30/40 | — | 70%28/40 |
| GPT 5.4 | 45%18/40 | — | 42%17/40 | 60%24/40 | 78%31/40 | 72%29/40 | 72%29/40 | 52%21/40 | 58%23/40 | — | 70%28/40 |
| Kimi K2.6 | 48%19/40 | 58%23/40 | — | 47%28/60 | 53%32/60 | 55%33/60 | 58%29/50 | 63%38/60 | 54%27/50 | 57%34/60 | 65%39/60 |
| GLM-5.1 | 38%15/40 | 40%16/40 | 53%32/60 | — | 37%22/60 | 62%37/60 | 50%25/50 | 67%40/60 | 68%34/50 | 73%44/60 | 58%35/60 |
| DeepSeek V4 Pro | 22%9/40 | 22%9/40 | 47%28/60 | 63%38/60 | — | 57%34/60 | 54%27/50 | 52%31/60 | 54%27/50 | 77%46/60 | 63%38/60 |
| GLM-5.2 | 45%18/40 | 28%11/40 | 45%27/60 | 38%23/60 | 43%26/60 | — | 55%22/40 | 67%40/60 | 75%30/40 | 55%33/60 | 60%36/60 |
| Claude Sonnet 4.6 | 32%13/40 | 28%11/40 | 42%21/50 | 50%25/50 | 46%23/50 | 45%18/40 | — | 48%19/40 | 65%26/40 | 70%14/20 | 68%27/40 |
| MiniMax-M3 | 28%11/40 | 48%19/40 | 37%22/60 | 33%20/60 | 48%29/60 | 33%20/60 | 52%21/40 | — | 40%16/40 | 50%30/60 | 48%29/60 |
| Claude Opus 4.8 | 25%10/40 | 42%17/40 | 46%23/50 | 32%16/50 | 46%23/50 | 25%10/40 | 35%14/40 | 60%24/40 | — | 50%10/20 | 45%18/40 |
| MiniMax-M2.7 | — | — | 43%26/60 | 27%16/60 | 23%14/60 | 45%27/60 | 30%6/20 | 50%30/60 | 45%9/20 | — | 65%39/60 |
| Qwen3.7 Plus | 30%12/40 | 30%12/40 | 35%21/60 | 42%25/60 | 37%22/60 | 40%24/60 | 32%13/40 | 50%30/60 | 55%22/40 | 33%20/60 | — |
2 · Why — what makes a model win or lose matches
How each model wins — and how it loses. Every match ends one of four ways; this bar
splits each model's matches into them, wins on the left, losses on the right (so the green→red
edge is its win rate):
bust-win — busted the opponent
cap-win — out-chipped them over 30 hands
cap-lose — behind on chips at the cap (ground down)
bust-out — busted out
A wide cap-win block = a grinder that out-chips
opponents; a wide cap-lose block = it gets
slowly ground down; a wide bust-out
block = it busts out a lot (high variance).
1.
2. GPT 5.4
61%match win
Strong — aggressive; fights back when behind.
bust-win 6%cap-win 55%cap-lose 27%bust-out 12%
3. Kimi K2.6
56%match win
Strong.
bust-win 11%cap-win 45%cap-lose 34%bust-out 10%
4. GLM-5.1
56%match win
Strong — fights back when behind.
bust-win 7%cap-win 48%cap-lose 36%bust-out 9%
5. DeepSeek V4 Pro
53%match win
Mid — steady (rarely busts out).
bust-win 6%cap-win 47%cap-lose 42%bust-out 4%
6. GLM-5.2
51%match win
Mid — steady (rarely busts out).
bust-win 10%cap-win 41%cap-lose 41%bust-out 8%
7. Claude Sonnet 4.6
48%match win
Mid — high-variance (busts out a lot); freezes when short (11%).
bust-win 14%cap-win 34%cap-lose 37%bust-out 15%
8. MiniMax-M3
42%match win
Weak.
bust-win 13%cap-win 29%cap-lose 48%bust-out 10%
9. Claude Opus 4.8
40%match win
Weak — freezes when short (6%).
bust-win 15%cap-win 26%cap-lose 48%bust-out 12%
10. MiniMax-M2.7
42%match win
Weak — high-variance (busts out a lot).
bust-win 11%cap-win 31%cap-lose 45%bust-out 12%
11. Qwen3.7 Plus
39%match win
Weak — very passive (rarely raises); steady (rarely busts out).
bust-win 7%cap-win 32%cap-lose 56%bust-out 5%
⚖️ Gear-shift: ahead vs behind (does it change how it plays when losing?)
Well sampled (every decision). Each cell is when ahead → when behind (change).
Aggression = (bet+raise+all-in) ÷ (bet+raise+all-in+call+check), folds excluded;
fold-to-bet = how often it folds when facing a bet.
Strong players shift gears when behind — they raise more (aggression ↑) and
fold less to bets (fold-to-bet ↓), i.e. they fight for pots; the change is
green when it shifts toward fighting,
red when it backs off. Weak players show
≈0 — they play the same whether winning or losing.
| model | aggression (ahead → behind) | fold-to-bet (ahead → behind) |
|---|---|---|
| GPT 5.5 | 34% → 51% (+17) | 24% → 16% (-8) |
| GPT 5.4 | 31% → 52% (+21) | 28% → 16% (-12) |
| Kimi K2.6 | 29% → 38% (+9) | 40% → 33% (-7) |
| GLM-5.1 | 20% → 32% (+12) | 38% → 32% (-6) |
| DeepSeek V4 Pro | 23% → 33% (+10) | 47% → 43% (-4) |
| GLM-5.2 | 23% → 32% (+9) | 33% → 26% (-7) |
| Claude Sonnet 4.6 | 29% → 32% (+3) | 30% → 29% (-1) |
| MiniMax-M3 | 26% → 28% (+2) | 30% → 25% (-5) |
| Claude Opus 4.8 | 25% → 25% (0) | 19% → 19% (0) |
| MiniMax-M2.7 | 30% → 37% (+7) | 30% → 25% (-5) |
| Qwen3.7 Plus | 14% → 16% (+2) | 34% → 34% (0) |
📊 Lead trajectory (each row = a model; 30 blocks = hands 1→30; block colour = share of matches it is ahead / behind on chips after that hand)
A row that stays green across =
wire-to-wire leader; green that reddens left→right = front-runs then gets ground down; green in the
middle that reddens at the end = builds a lead but
can't close it; all red = behind the whole match.
The last block ≈ the model's match win rate.
behind ≤35%ahead ≥65%
50% = even
hand →151015202530
GPT 5.565%GPT 5.461%Kimi K2.656%GLM-5.156%DeepSeek V4 Pro53%GLM-5.251%Claude Sonnet 4.648%MiniMax-M342%Claude Opus 4.840%MiniMax-M2.742%Qwen3.7 Plus39%3 · Analysis
A 30-hand match is usually won by leading at the hand cap, not by busting the
opponent — so the edge is chip management: fighting back when behind, gear-changing as stacks
get short, and not blowing up a whole match in one hand. Note a real limitation of this fixed
dataset: the strong models rarely reach a short stack at all (they protect their chips), so
their short-stack columns are based on very few decisions — greyed cells (n<30) are
suggestive, not conclusive. We can't add more matches, so we read those honestly.
🪜 Aggression by stack depth (does it push/fold when short?)
Aggression = (bet+raise+all-in) ÷ (bet+raise+all-in+call+check), folds
excluded — bucketed by effective stack:
deep ≥40bb · mid 15–40bb · short <15bb (push-fold territory). Greener =
more aggressive; greyed cells have n<30 (too few decisions — read as suggestive). The
n on the short column doubles as an exposure signal: the top models barely appear there
because they rarely get short-stacked.
| model | deep (≥40bb) | mid | short (<15bb) |
|---|---|---|---|
| GPT 5.5 | 40% n=23054 | 37% n=494 | 86% n=14 |
| GPT 5.4 | 40% n=24339 | 32% n=697 | 24% n=41 |
| Kimi K2.6 | 33% n=29639 | 24% n=755 | 30% n=76 |
| GLM-5.1 | 27% n=32978 | 25% n=766 | 50% n=12 |
| DeepSeek V4 Pro | 28% n=29410 | 21% n=334 | 46% n=13 |
| GLM-5.2 | 28% n=34003 | 26% n=507 | 18% n=55 |
| Claude Sonnet 4.6 | 31% n=24456 | 24% n=891 | 11% n=65 |
| MiniMax-M3 | 28% n=32922 | 22% n=998 | 35% n=69 |
| Claude Opus 4.8 | 26% n=28430 | 18% n=1284 | 6% n=81 |
| MiniMax-M2.7 | 34% n=24474 | 35% n=553 | 33% n=33 |
| Qwen3.7 Plus | 15% n=35614 | 12% n=782 | 0% n=28 |
🧪 Decision quality (experimental — uses each model's hole-card equity vs a random hand; the opponent's range is never known, so read this as a gross, range-free signal, not exact EV)
Does it bluff, or only bet strong hands?
bluff rate = share of its bets/raises made with weak cards (<40% equity);
bluff success = of those bluffs, how often the opponent actually folds; avg equity when
betting = how strong its cards are, on average, when it fires. A model that only bets strong
hands (low bluff rate, high avg equity) is predictable and easy to fold to; mixing in
bluffs is a sign of skill.
corr with win%: bluff rate +0.61
(more bluffing ↔ winning), avg-equity-when-betting -0.64
(only-bet-strong ↔ losing); bluff success -0.12
(not a skill signal).
| model | bluff rate (<40% eq) | bluff success | avg equity when betting | n |
|---|---|---|---|---|
| GPT 5.5 | 9% | 46% | 58% | 4471 |
| GPT 5.4 | 10% | 38% | 56% | 5180 |
| Kimi K2.6 | 4% | 36% | 59% | 6675 |
| GLM-5.1 | 4% | 41% | 63% | 4122 |
| DeepSeek V4 Pro | 4% | 41% | 61% | 4908 |
| GLM-5.2 | 4% | 55% | 65% | 3665 |
| Claude Sonnet 4.6 | 2% | 59% | 65% | 3004 |
| MiniMax-M3 | 4% | 42% | 61% | 4993 |
| Claude Opus 4.8 | 1% | 46% | 65% | 3490 |
| MiniMax-M2.7 | 9% | 46% | 60% | 3344 |
| Qwen3.7 Plus | 1% | 37% | 66% | 2912 |
🎭 Player behaviour profiles
| # | model | VPIP | PFR | aggr | fold→bet | bet size | all-in% | WTSD | W@SD | hand-win | tokens/dec |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | GPT 5.5 | 76% | 38% | 40% | 21% | 1.09x | 1% | 27% | 45% | 59% | 294 |
| 2 | GPT 5.4 | 75% | 44% | 40% | 23% | 1.02x | 0% | 31% | 48% | 55% | 905 |
| 3 | Kimi K2.6 | 60% | 43% | 33% | 37% | 1.25x | 0% | 29% | 52% | 41% | 7,166 |
| 4 | GLM-5.1 | 61% | 21% | 27% | 36% | 1.06x | 0% | 29% | 50% | 40% | 3,056 |
| 5 | DeepSeek V4 Pro | 51% | 28% | 28% | 45% | 1.25x | 0% | 25% | 57% | 37% | 1,159 |
| 6 | GLM-5.2 | 68% | 17% | 28% | 29% | 1.19x | 0% | 31% | 42% | 44% | 1,454 |
| 7 | Claude Sonnet 4.6 | 68% | 20% | 30% | 29% | 1.01x | 1% | 29% | 38% | 46% | 157 |
| 8 | MiniMax-M3 | 61% | 30% | 28% | 28% | 1.26x | 0% | 35% | 43% | 44% | 2,236 |
| 9 | Claude Opus 4.8 | 74% | 26% | 25% | 19% | 1.10x | 1% | 42% | 39% | 48% | 55 |
| 10 | MiniMax-M2.7 | 63% | 22% | 34% | 28% | 1.09x | 1% | 30% | 41% | 48% | 6,027 |
| 11 | Qwen3.7 Plus | 64% | 17% | 15% | 34% | 1.10x | 0% | 37% | 56% | 36% | 1,166 |
VPIP = voluntarily entered the pot preflop (looseness). PFR = preflop raise. aggr = aggression = (bet+raise+all-in) ÷ (bet+raise+all-in+call+check); folds are excluded. fold→bet = folds when facing a bet. bet size = avg bet/raise as a multiple of the pot. WTSD = went to showdown; W@SD = won at showdown. hand-win = share of hands that netted chips. tokens/dec = avg reasoning tokens per decision.
🎭 Style map (VPIP vs aggression — top-right = loose & aggressive; aggression = (bet+raise+all-in) ÷ (bet+raise+all-in+call+check), folds excluded)
🧮 Thinking effort (avg reasoning tokens / decision)
🃏 Action mix (% of all decisions)
📈 Aggression by street (aggression = (bet+raise+all-in) ÷ (bet+raise+all-in+call+check), folds excluded; darker = more aggressive)
preflopflopturnriver
GPT 5.538474229GPT 5.445424025Kimi K2.654252214GLM-5.125292823DeepSeek V4 Pro38262415GLM-5.219373324Claude Sonnet 4.621393727MiniMax-M336252518Claude Opus 4.826312120MiniMax-M2.725403933Qwen3.7 Plus19141311