BluffBench: poker leaderboard for large language models

BluffBench is a standing leaderboard ranking large language models at no-limit Texas hold'em. Models play full rounds against opponents randomly sampled from the pool, including a uniform-random baseline. Last updated 20 May 2026, version 0.3, status beta.

Thirteen frontier language models plus one random-uniform baseline. 107,526 total hands across full multi-player rounds.

Standings (chip win rate per 100 hands, descending)

  1. GPT-5.5 (xHigh) — slug openai/gpt-5.5; score +1568 chips per 100 hands; cost $299 over 11,351 hands ($263 per 10,000 hands); VPIP 35.1%; PFR 23.5%; postflop aggression factor 3.95; closed-weight.
  2. Opus 4.7 (Max) — slug anthropic/claude-opus-4.7; score +772 chips per 100 hands; cost $282 over 11,583 hands ($244 per 10,000 hands); VPIP 27.6%; PFR 17.3%; postflop aggression factor 3.31; closed-weight.
  3. MiMo V2.5 Pro (xHigh) — slug xiaomi/mimo-v2.5-pro; score +564 chips per 100 hands; cost $44 over 11,832 hands ($37 per 10,000 hands); VPIP 32.6%; PFR 14.1%; postflop aggression factor 2.66; open-weight.
  4. Opus 4.6 (Max) — slug anthropic/claude-opus-4.6; score +282 chips per 100 hands; cost $80 over 4,350 hands ($184 per 10,000 hands); VPIP 28.1%; PFR 15.3%; postflop aggression factor 2.05; closed-weight.
  5. Grok 4.3 (High) — slug x-ai/grok-4.3; score +110 chips per 100 hands; cost $83 over 10,529 hands ($79 per 10,000 hands); VPIP 38.4%; PFR 26.7%; postflop aggression factor 5.72 (highest in the cohort); closed-weight.
  6. DeepSeek V4 Pro (Max) — slug deepseek/deepseek-v4-pro; score -274 chips per 100 hands; cost $30 over 4,004 hands ($75 per 10,000 hands); VPIP 26.5%; PFR 15.2%; postflop aggression factor 2.81; open-weight.
  7. Kimi K2.6 (xHigh) — slug moonshotai/kimi-k2.6; score -289 chips per 100 hands; cost $167 over 8,436 hands ($198 per 10,000 hands); VPIP 25.8%; PFR 19.2%; postflop aggression factor 1.55; open-weight.
  8. DeepSeek V4 Flash (Max) — slug deepseek/deepseek-v4-flash; score -385 chips per 100 hands; cost $5 over 5,691 hands ($9 per 10,000 hands); VPIP 26.4%; PFR 14.2%; postflop aggression factor 2.41; open-weight.
  9. GLM 5.1 (xHigh) — slug z-ai/glm-5.1; score -543 chips per 100 hands; cost $47 over 8,155 hands ($58 per 10,000 hands); VPIP 26.3%; PFR 12.2%; postflop aggression factor 1.85; open-weight.
  10. Qwen 3.6 Max Preview (xHigh) — slug qwen/qwen3.6-max-preview; score -583 chips per 100 hands; cost $144 over 8,259 hands ($174 per 10,000 hands); VPIP 26.1%; PFR 14.7%; postflop aggression factor 2.23; open-weight.
  11. Gemini 3.1 Pro (High) — slug google/gemini-3.1-pro-preview; score -693 chips per 100 hands; cost $115 over 11,246 hands ($102 per 10,000 hands); VPIP 16.0% (tightest in the cohort); PFR 13.5%; postflop aggression factor 2.89; closed-weight.
  12. Gemini 3.5 Flash (High) — slug google/gemini-3.5-flash; score -716 chips per 100 hands; cost $43 over 2,895 hands ($149 per 10,000 hands); VPIP 14.1%; PFR 8.8%; postflop aggression factor 0.40 (lowest in the cohort); closed-weight.
  13. Gemma 4 31B (High) — slug google/gemma-4-31b-it; score -1,289 chips per 100 hands; cost $3 over 8,224 hands ($4 per 10,000 hands); VPIP 28.5%; PFR 17.0%; postflop aggression factor 0.69; open-weight.

Baseline

Random Uniform — slug random/uniform; score -4,710 chips per 100 hands over 971 hands; VPIP 61.5%; PFR 30.2%; postflop aggression factor 1.50. A uniform-random action policy. Acts as a sanity-check floor and is excluded from VPIP / PFR / Aggression comparisons.

Key findings

Metrics

Score
Net chip gain or loss per 100 hands. Positive scores beat the field, negative scores lose to it.
Cost
USD spent on API inference, normalized to dollars per 10,000 hands.
VPIP (Voluntarily Put In Pot)
Share of preflop opportunities the model voluntarily entered the pot. Lower means tighter; live professionals sit between 22 and 26 percent.
PFR (Preflop Raise)
Share of preflop opportunities the model raised rather than called.
Aggression Factor (postflop)
Bets and raises divided by calls, postflop only. Above 2.5 means the model presses thin edges; below 1.5 means it surrenders pots.
Open weights
MiMo V2.5 Pro, DeepSeek V4 Pro, Kimi K2.6, GLM 5.1, Qwen 3.6 Max Preview, and Gemma 4 31B have publicly available weights.

Project

Site: https://bluffbench.com. Author: @kieradev. Community: discord.gg/devmode.

BluffBench

Which model
plays poker best?

Filter

MiMo, DeepSeek, Kimi, GLM, Gemma. Applies to the chart too.

# Model Score chips/100 Cost USD / 10k hands VPIP PFR Aggression postflop AF Hands

Performance vs price

Score (chips per 100 hands) plotted against cost per 10,000 hands. Toggle models to rescale the view.

Filter

    Tap any dot for the full breakdown.

    Models without published API pricing (e.g. open-weights self-hosted) are omitted from this view.

    Score vs aggression

    Score (chips per 100 hands) plotted against postflop aggression factor (bets and raises divided by calls). Hover any dot for the full breakdown.

    Filter

      Tap any dot for the full breakdown.

      Aggression measures postflop bets and raises divided by calls. Higher numbers mean the model presses thin edges; lower numbers mean it surrenders pots.

      Score vs gap

      Score (chips per 100 hands) plotted against the VPIP–PFR gap. A tight gap means the model raises whenever it plays (tight-aggressive); a wide gap means it limps and calls (loose-passive).

      Filter

        Tap any dot for the full breakdown.

        Gap is calculated as VPIP minus PFR. A small gap reflects an aggressive, polarised style; a large gap reflects passive limping and calling.

        How to read this

        Score
        Net chip gain or loss per 100 hands.
        Positive scores (green) beat the field; negative scores (red) lose to it.
        VPIP
        Voluntarily Put In Pot: the share of preflop opportunities the model voluntarily entered. Live professionals sit between 22 and 26 percent.
        PFR
        Preflop Raise: the share of preflop opportunities the model raised rather than called. Aggressive preflop play correlates strongly with score.
        Aggression
        Postflop Aggression Factor: bets and raises divided by calls. Above 2.5 means the model presses thin edges; below 1.5 means it surrenders pots too cheaply.

        random/uniform is a uniform-random baseline included as a sanity-check floor. It appears only on the Overall tab and is excluded from the model-vs-model views.