chrystians/poker-holdembench

HoldemBench: 17 AIs, 7,700 hands of heads-up no-limit hold'em

Every hand and every decision from HoldemBench, a benchmark that asks which kind of AI makes better sequential decisions under uncertainty: OpenRouter decision models (Decisions API), cheap reasoning LLMs, frontier LLMs, a local open-weights model, and plain scripts.

All games are heads-up no-limit Texas hold'em, blinds 50/100, 100 big blinds per player, stacks reset every hand. Every deal is played twice with the seats swapped (duplicate poker), and all-in hands before the river are scored by their exact equity, so most card luck cancels out.

Players

Tier	Players
Decision models (Decisions API)	Jev 1.13, pplx-decider 27B, Solar Decide, Mercury Decide
Cheap reasoning LLMs	GPT-6 Luna, Muse Spark 1.3 Contributor, DeepSeek V4.1 Flash, GLM 5.3 Flash, MiMo V2.6 Flash
Frontier LLMs	GPT-6.1 Sol, Claude Opus 5.5 (default reasoning)
Local open-weights	Bielik 11B v3.0 Instruct (Q8_0 GGUF, llama.cpp)
Scripts	equity_tag (Monte Carlo equity, ~30 lines), rock, maniac, calling station, random

Files

File	Rows	Content
decisions.jsonl	~52,000	One row per decision: private cards, board, pot, stack, legal menu, chosen action, the player's own win-probability estimate, the full option distribution (decision models), model id, latency, cost in USD, output tokens, the hand's final chip result.
hands.jsonl	~9,700	Full hand records: both players' cards, the five board cards, every action with its metadata, payoffs, all-in-adjusted payoffs, showdown flag. Includes the bot-only calibration league.
probes.jsonl	~2,000	40 frozen spots asked to every player three times: original, repeated, and with suits renamed (strategically identical).
spots.json	40	The frozen probe spots with their full prompt text.
leaderboard.json	17	Final ratings (BB/hand, all-in adjusted, with 95% bootstrap intervals), Elo on 10-deal blocks, cost and latency per player.

decisions.jsonl columns

Column	Type	Meaning
run, match_id, hand_index, deal_seed	str / int	League, pairing, hand within the match, deal seed (both orientations of a deal share it)
decision	int	Index of the action within the hand
agent, opponent, seat	str	Acting player, opponent, SB or BB
street, hole, board, pot, to_call, stack		Situation the player saw (chips; big blind = 100)
options	list[str]	Legal menu, e.g. fold, call, raise_400, all_in (numbers are chip totals)
action	str	Chosen option
fallback	bool	True when the model's answer was unusable and check/fold was played instead
equity_estimate	float or null	The player's own stated chance to win (0–1)
probabilities	dict or null	Decision models only: probability of every option
model, latency_s, cost_usd, output_tokens		Model id, wall time, provider-reported cost and output tokens of the call
hand_result_chips	int	The player's chip result for the whole hand (not all-in adjusted)

hands.jsonl has one row per hand: players, seats, holes, full_board, actions (each with pot_before, to_call, added, equity_est, probabilities, fallback), payoffs, allin_adjusted (all-in hands scored by equity, the benchmark's primary result), showdown, deal_index and orientation. leaderboard.json is written by the same analysis as the technical report and the write-up, so its numbers match them exactly.

run names the league: run1a-decision-league (decision models and bots), run1a-mercury, run1b-chat-league (cheap LLMs), run1c-bielik (local model), run2-frontier and run2b-confirm (frontier models), bots (bot-only calibration).

Headline results

• Cheap reasoning LLMs (DeepSeek V4.1 Flash +5.4, Muse Spark Contributor +4.3, Luna +3.0 BB/hand) played as well as GPT-6.1 Sol (+3.4) and Claude Opus 5.5 (+2.7), at 1/10 to 1/30 of the price.
• The ~30-line equity_tag script beat every decision model it played enough hands against (Jev by 2.7 BB/hand over 100 duplicate deals) and drew with Opus over 70 deals.
• Decision models are deterministic but play exploitable poker (Jev's median postflop bet is 6x pot; Solar goes all-in in 56% of hands and folded aces).
• Bielik finished last (−8.4 BB/hand): it never folded to a bet.

Notes

• Hidden reasoning is not included; prompts can be rebuilt exactly from the state with holdembench.engine.Observation.render.
• Hands voided by provider failures (rate limits, timeouts, HTTP errors) were replayed and are excluded; model failures (truncated or unparseable answers) count against the model and appear as fallback: true.
• Model outputs are subject to each provider's terms of use.

Code, methodology and technical report: https://github.com/ChrystianSchutz/poker-holdembench · Write-up: https://chrystianschutz.com/blog/ai-poker-holdembench/

Citation

@techreport{schutz2026holdembench,
  title  = {HoldemBench: Duplicate Heads-Up Hold'em as a Cheap, Variance-Controlled Benchmark
            for Decision Models, Chat LLMs, Local Models and Scripts},
  author = {Schutz, Chrystian},
  year   = {2026},
  month  = oct,
  url    = {https://chrystianschutz.com/blog/ai-poker-holdembench/}
}