fp.

PokerBench is my attempt at a new LLM benchmark wherein frontier models play Texas Hold'em in an arena setting. It also features a simulator to view individual games and observe how the different models reason about poker strategy. Opus/Haiku, Gemini Pro/Flash, GPT-5.2/5 mini, and Grok 4.1 Fast Reasoning have all been included.

All code -> https://github.com/JoeAzar/pokerbench

Comments

tcpais•4w ago

Finally, a way to settle the model wars that actually matters: Texas Hold'em. That 3D replay view is sick! ♠♦ I spent way too long watching the replay on Game 2a58900d. It’s wild to see the chain of thought mapped against the betting rounds. It really exposes when a model is hallucinating a strong hand versus actually calculating pot odds. This 'PokerBench' might actually become the standard for measuring agentic risk-taking.

falloutx•4w ago

yeah the 3d view is amazing

VK-pro•4w ago

Very very fun. Just glancing at this quickly at lunch but is there any idea of incorporating tool use?

jazarwil•4w ago

Not at the moment, do you have something in mind?

thorawaytrav•4w ago

Do you have idea why smaller models are better then large ones?

jazarwil•4w ago

I've seen some theories tossed around but I don't think I'm qualified to offer an authoritative answer. Gemini 3 Pro specifically seems to be consistently "tighter" and more passive than Flash.

falloutx•4w ago

Fun, any idea how much would be the cost per game? I am worried 160 isnt a big enough sample size.

jazarwil•4w ago

It greatly depends on the models. The 6-handed setup with Opus and Pro cost about $30/game. The 4-handed setup with just small models was $6/game. I'd love to run more but I already spent quite a bit as it is.

falloutx•4w ago

Yeah thats costly, 160 games still gives about 1000+ total decisions and you can see some trends on how they think about the game state.

jazarwil•4w ago

Oh to be clear, there are ~21k hands here, and far more decisions than that.

Onavo•4w ago

What about the open source models? I remember from the trading benchmarks Deepseek performed pretty well.

jazarwil•4w ago

I didn't incorporate any open weights/source models just to limit the number of API providers I had to juggle, but it is just a config change if somebody wants to try a run with them.

alalani1•4w ago

Do you have any idea why the win rate for GPT-5.2 is higher than Gemini 3 Flash yet the former loses money while the latter earns money? Is it just bet sizing (betting more when it has a good hand) or something else?

jazarwil•4w ago

There are a few reasons that come to mind, such as winning larger pots on average, and also playing more hands by virtue of not getting knocked out as frequently.

tanvach•4w ago

People looking into this a little too much, looks to me like random walk. You should try reinitiating the trial (or have multiple running) and see if the ranking is robust.

jazarwil•4w ago

Wdym exactly? I ran 163 games, are you suggesting more games or something else?

whattheheckheck•4w ago

You need to simulate 50k to 200k hands to get a true winrate

jazarwil•4w ago

I'd love to run more games, just very expensive unfortunately.

alfonsodev•4w ago

Really cool, I’m curious what would be the comparison versus a deterministic bot that uses probability tables.