GTO Wizard AI is a modern poker-specific AI agent developed to power the GTO Wizard platform. The technology originally launched under the name Ruse AI, created in Canada by Marc-Antoine Provost and Philippe Beardsell, before being integrated into GTO Wizard in 2023.
Unlike earlier-generation poker bots—such as the 2018 ACPC champion Slumbot—this model does not rely on precomputed strategies. Instead, it:
- trained against itself over hundreds of millions of hands,
- continuously optimized decisions based on expected value (EV),
- and solves each situation in real time.
This approach is based on deep reinforcement learning, allowing the AI to determine optimal decisions within seconds for any given scenario.
Dominant Performance: Results vs Slumbot
The capabilities of GTO Wizard AI were tested in a 150,000-hand heads-up match against Slumbot.
Result: +19.4 bb/100 win rate
This is an exceptionally strong outcome. At $50/$100 stakes, this translates to approximately $19.4 profit per hand on average over 100 hands, which—depending on game speed—can scale to several thousand dollars per hour.
General AI vs Poker-Specific Models
The following results come from one of the first comprehensive benchmarks directly comparing:
- large language models (LLMs), and
- poker-optimized AI systems
The conclusions are clear:
- General models perform well in reasoning and language tasks,
- but they remain significantly behind poker-specific systems in strategic decision-making.
It is worth noting, however, that GPT models already outperform earlier-generation bots such as Slumbot.
GTO Wizard AI Results vs Other Models
- GPT-5.3 – +16 bb/100
- GPT-5.4 – +17.8 bb/100
- Claude Opus 4.6 – +20.4 bb/100
- Gemini 3.1 Pro – +30.8 bb/100
- Grok 4 – +60 bb/100
Removing Variance: How AIVAT Works
One of poker’s defining characteristics is variance—the fact that luck can heavily influence short-term results. Traditionally, hundreds of thousands of hands are required to draw reliable conclusions.
This is where AIVAT (Action-Informed Value Assessment Tool) comes in.
AIVAT assigns a theoretical “expected” outcome to each decision point and compares it to the actual result. Rather than measuring profit alone, it evaluates how much a player or model deviates from the optimal expected value in each situation.
As a result:
- the impact of luck is significantly reduced,
- reliable conclusions can be drawn from far fewer hands,
- and a variance-adjusted win rate can be calculated.
This shifts the focus from short-term outcomes to decision quality.
Conclusion
The data paints a clear picture:
General-purpose AI models are not yet able to compete with systems specifically designed for poker—and the gap is substantial.
Current trends suggest that specialized AI will continue to hold a strong advantage in environments where decision-making requires precise, mathematically grounded optimization.
0 comments