$ ~/aibattle/leduc

🎴 Leduc Holdem · Imperfect-information poker · 6-card deck · round-robin, seat-swapped · 3850 games

🎬 Watch featured replays → Setup & rules · expand

How Leduc Holdem works

A tiny imperfect-information poker on a 6-card deck — two each of JQK (J < Q < K). Heads-up, two betting rounds:

Both players ante 1 chip, then each is dealt one private card.
Round 1 (pre-flop): players bet on their private card alone. Actions are check/bet and call/raise/fold; bets this round are 2 chips.
One public card is then revealed (the “flop”), shared by both.
Round 2: another betting round, now with bets of 4 chips.
At showdown, pairing your private card with the public card wins; otherwise the higher card wins. A fold concedes the pot.

Small enough to reason about precisely, yet it has genuine bluffing, value-betting and pot odds — pairing the board is the nuts, and a lone King bluffs well. Rated by a chip-weighted Elo (like Hold'em), so the size of pots won/lost matters, not just who won the hand.

What the model sees each turn: its own private card, the public card once it is revealed, the pot and the bets so far, and its legal actions — never the opponent's card.

🏆 Leaderboard

#	model	Elo	net/game	win%	draw%	1st-move win%	invalid%	plies	think	games	tokens/dec	$/1K dec
1	GPT 5.4	1511 ±21	+0.07	45%	18%	41%	0.0%	4.3	58.4s	450	—	—
2	GLM-5.2	1511 ±17	+0.06	41%	20%	40%	0.0%	4.0	60.9s	800	3,013	$13.26
3	Kimi K2.6	1506 ±16	+0.04	40%	17%	40%	0.0%	3.9	238.3s	850	11,037	$44.15
4	Qwen3.7 Plus	1506 ±18	+0.02	39%	17%	38%	0.0%	4.0	25.0s	800	2,124	$3.40
5	Claude Sonnet 4.6	1503 ±22	+0.01	44%	16%	44%	0.0%	3.9	3.1s	450	—	—
6	GPT 5.5	1503 ±24	+0.00	40%	19%	41%	0.0%	3.9	13.9s	450	—	—
7	MiniMax-M3	1502 ±18	+0.00	42%	16%	44%	0.0%	3.9	38.2s	850	2,743	$3.29
8	GPT-OSS 120B	1501 ±20	-0.00	44%	14%	45%	0.0%	3.9	8.4s	650	1,381	$0.83
9	GLM-5.1	1500 ±17	-0.01	40%	21%	36%	0.0%	4.1	24.5s	850	1,618	$7.12
10	DeepSeek V4 Pro	1498 ±16	-0.02	39%	19%	40%	0.0%	4.1	104.2s	850	1,882	$6.55
11	Claude Opus 4.8	1484 ±26	-0.13	39%	21%	43%	0.0%	3.9	1.9s	450	—	—
12	MiniMax-M2.7	1476 ±32	-0.18	42%	15%	40%	0.0%	4.0	81.5s	250	16,247	$19.50

columns: Elo opponent-adjusted rating · net/game mean result/game (+1/−1/0) · win% / draw% won / drawn · 1st-move win% win rate acting first · invalid% illegal moves · plies avg game length · think avg sec/decision · games games played (varies by wave)

Elo rating

Win / draw / loss

Net result per game

Game-length distribution (plies)

⚔️ Head-to-head (row wins–losses vs column)

	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-M3	Qwen3.7 Plus	GPT-OSS 120B	MiniMax-M2.7
Claude Opus 4.8	—	21-24 5d	20-20 10d	19-23 8d	18-17 15d	18-22 10d	18-19 13d	19-18 13d	19-14 17d	22-23 5d	0-0 0d	0-0 0d
Claude Sonnet 4.6	24-21 5d	—	21-18 11d	18-20 12d	22-22 6d	17-24 9d	24-24 2d	21-17 12d	23-20 7d	27-16 7d	0-0 0d	0-0 0d
DeepSeek V4 Pro	20-20 10d	18-21 11d	—	36-40 24d	36-40 24d	19-25 6d	20-22 8d	36-38 26d	42-43 15d	45-37 18d	41-44 15d	21-21 8d
GLM-5.1	23-19 8d	20-18 12d	40-36 24d	—	36-40 24d	19-26 5d	21-20 9d	40-38 22d	39-40 21d	42-33 25d	42-43 15d	18-16 16d
GLM-5.2	17-18 15d	22-22 6d	40-36 24d	40-36 24d	—	18-18 14d	24-16 10d	41-41 18d	46-40 14d	41-39 20d	39-42 19d	0-0 0d
GPT 5.4	22-18 10d	24-17 9d	25-19 6d	26-19 5d	18-18 14d	—	19-17 14d	22-18 10d	21-22 7d	25-21 4d	0-0 0d	0-0 0d
GPT 5.5	19-18 13d	24-24 2d	22-20 8d	20-21 9d	16-24 10d	17-19 14d	—	21-22 7d	21-18 11d	22-17 11d	0-0 0d	0-0 0d
Kimi K2.6	18-19 13d	17-21 12d	38-36 26d	38-40 22d	41-41 18d	18-22 10d	22-21 7d	—	43-45 12d	43-45 12d	43-47 10d	23-25 2d
MiniMax-M3	14-19 17d	20-23 7d	43-42 15d	40-39 21d	40-46 14d	22-21 7d	18-21 11d	45-43 12d	—	45-40 15d	47-44 9d	23-21 6d
Qwen3.7 Plus	23-22 5d	16-27 7d	37-45 18d	33-42 25d	39-41 20d	21-25 4d	17-22 11d	45-43 12d	40-45 15d	—	38-44 18d	0-0 0d
GPT-OSS 120B	0-0 0d	0-0 0d	44-41 15d	43-42 15d	42-39 19d	0-0 0d	0-0 0d	47-43 10d	44-47 9d	44-38 18d	—	24-21 5d
MiniMax-M2.7	0-0 0d	0-0 0d	21-21 8d	16-18 16d	0-0 0d	0-0 0d	0-0 0d	25-23 2d	21-23 6d	0-0 0d	21-24 5d	—