AI researchers have long loved poker as a test for one reason: you don't get to see everything. In chess or Go both players see the whole board, so a program can compute. In hold'em you don't see your opponent's cards, bets can be bluffs, and you have to vary your play so nobody can read it. Between 2015 and 2019, academic groups went from solving a small but serious version of the game to beating pros at a six-player table. Then the questions changed. Can general chatbots play? Are the bots you can actually buy as good as their marketing? Here are the big studies, the numbers they reported, and what those numbers do and don't mean.How poker AI results are measuredMost poker AI papers give win rates in milli-big-blinds per game (mbb/g), meaning thousandths of a big blind won per hand. Players usually think in big blinds per 100 hands (bb/100). So 10 mbb/g = 1 bb/100. Between human pros, 50 mbb/g (5 bb/100) is already considered a large edge. Poker is a noisy game, and a few thousand hands can produce nearly any outcome by chance. That is why serious studies report confidence intervals or standard errors, and many use a variance-reduction technique called AIVAT that takes a lot of the luck out of the estimate.The milestones in one tableYearSystem or studyGameResult2015Cepheus (University of Alberta)Heads-up limit hold'emEssentially solved: exploitability 0.986 mbb/g2015Claudico (Carnegie Mellon)Heads-up no-limitLost to four pros over 80,000 hands, statistically inconclusive2017DeepStack (University of Alberta)Heads-up no-limit486 mbb/g against professionals, 44,852 hands2017Libratus (Carnegie Mellon)Heads-up no-limit147 mbb/g against four top specialists, 120,000 hands2019Pluribus (Carnegie Mellon and Facebook AI)Six-player no-limit48 mbb/g against elite pros, 10,000 hands2023Student of Games (DeepMind)Chess, Go and heads-up no-limitOne algorithm; beat Slumbot, the strongest openly available poker agent2025PokerBenchLarge language modelsBest general model, GPT-4, matched the optimal action and bet size in 53.55% of spots2026Commercial bots compared (preprint)6-max and 9-max no-limit, bot vs botAI bots +29.2 bb/100, rule-based bots −24.3 bb/100The numbers in the last column can't be compared with each other. They come from different games, against different opponents, measured in different ways. Read each row as the answer to its own question.2015: limit hold'em is solved, no-limit is notScience published Cepheus in January 2015. It was the first program to essentially solve a poker game people actually play: heads-up limit hold'em. Its strategy is so close to the game-theoretic optimum that, according to the authors, a human lifetime of play wouldn't be enough to tell the difference. Getting there took an algorithm called CFR+ running on 4,800 CPUs for 68 days, about 900 core-years.That same spring, Carnegie Mellon's Claudico played four pros over 80,000 hands of heads-up no-limit. The humans came out ahead by $732,713 in virtual chips, but the organizers called it a statistical tie because the sample wasn't large enough to separate skill from variance. With free bet sizing the game tree is massive, and the brute-force approach that worked for limit didn't cut it.2017: two programs beat professionals at heads-up no-limitTwo groups got over the line within weeks of each other. The University of Alberta's DeepStack played 44,852 hands against 33 professionals from 17 countries. Eleven of them finished the full 3,000-hand match, and DeepStack beat ten of those eleven by a statistically significant margin. After the authors adjusted for luck, its win rate was 486 mbb/g, close to ten times what counts as a large margin between pros. The trick was a change of approach: instead of precomputing the entire game, DeepStack re-solved each situation on the fly and used a neural network to estimate the value of positions deeper in the hand.Libratus went the other way. In January 2017 it played four of the best heads-up specialists for 20 days and 120,000 hands. It won 147 mbb/g with 99.98% statistical significance and finished about $1.8 million ahead in chips. Its designers combined a precomputed blueprint strategy with solving later streets on the fly. Each night, the program patched the gaps in its strategy that the humans had found during the day.2019: the six-player table fallsHeads-up poker has a useful property: an exact equilibrium strategy can't lose in expectation (before rake, averaged over both seats). With five opponents that guarantee goes away, which made multiplayer poker a much harder problem. Pluribus, published in Science in 2019, played 10,000 hands of six-player no-limit hold'em over 12 days against elite pros, five of them at the table at a time. It won 48 mbb/g (roughly 4.8 bb/100) with a standard error of 25 mbb/g. In a second format, one human against five copies of the bot, it won 32 mbb/g.The more striking number is the cost. Pluribus learned its strategy by playing against itself for eight days on a single 64-core server, and the authors put the compute at under $150 at cloud prices. Compare that with about 900 core-years for Cepheus four years earlier. That drop in cost helped poker AI move out of university labs and into consumer products like PokerBotAI.2023: one algorithm for chess, Go and pokerIn November 2023, Science Advances published DeepMind's Student of Games. The paper asked whether one algorithm can handle both full-information and hidden-information games. Student of Games combines search, self-play and game theory. It reached strong play in chess and Go and beat Slumbot, the best openly available heads-up no-limit agent, by a slim margin of 7 mbb/g. It is weaker in chess and Go than specialized systems like AlphaZero, but it showed that tools built for poker carry over to other games.2025: can a chatbot play poker?When large language models arrived, many people expected a general AI to be a good poker player too. PokerBench, presented at AAAI 2025, put that to the test with 11,000 decision spots built with the help of trained poker players: 1,000 preflop and 10,000 postflop. The best pre-trained model was GPT-4. It matched the optimal action and bet size in 53.55% of spots, or 65.54% when only the action was counted. Other models did worse. Fine-tuning on poker data improved results a lot. The takeaway still stands: a model trained on text doesn't think like a solver, and it doesn't become one just because it can explain poker terms well.2026: what happens when commercial bots play each otherThe academic milestones have a lot in common. They were built by research teams on dedicated hardware, and most were tested against human pros or other research agents. They say little about the bots sold to the public, whose vendors often claim AI, GTO or machine learning without publishing long-term results.A preprint posted in October 2026 checks those claims directly. At the end of 2025, its author bought six commercial poker bots and had them play each other in seven configurations in a private PPPoker club, with no humans and no chips of cash value. The paper, a 26,395-hand bot-versus-bot comparison on Zenodo, hasn't been peer-reviewed, but it comes with the full hand history archive, so you can recount the numbers yourself.The products fall into two camps. Two bots used machine-learning decision engines. Four configurations were rule-based bots driven by player profiles, the older design built on hand-written if-then logic. One more product lost chips to everyone at random and was left out of the main analysis, which left 17,295 hands. In those hands:The AI group won +29.2 bb/100 (95% confidence interval: +16.9 to +43.1).The rule-based group lost −24.3 bb/100, with an interval of −33.5 to −16.7.The two intervals don't overlap each other or zero. Neither group changed sign in any slice the author looked at, including table size and date.A paid top-tier profile didn't save the rule-based approach: it ended up at −28.0 bb/100.The difference between the two AI bots stayed inside the margin of error, so the study doesn't pick a winner between them.The author is clear about the limits. This is bot versus bot at micro stakes, and it shouldn't be read as a prediction of how any of these programs would do against people. With only 30 table sessions, the intervals may be a little too narrow. And since the AI bots played rule-based opponents more often, the table mix affects the size of the gap more than its direction. Within those limits, the data shows a consistent difference between two generations of design.What the decade adds up toThree trends run through all of this. The compute bill went from hundreds of core-years to a few days on one machine. The games went from heads-up limit to six-player no-limit. And the methods moved from academic labs to products anyone can buy.The questions have changed too. Nobody seriously doubts anymore that a program can beat a top professional at a single table. The open questions are about measurement. Which vendor claims hold up over tens of thousands of hands? How do general AI models compare with programs built for one game? On both, published data with confidence intervals is still rare. Ask for it before you trust a headline.FAQHas AI solved poker?The only popular form of poker that has essentially been solved is heads-up limit hold'em, by the Cepheus program in 2015. Its strategy is so close to the theoretical optimum that you could play it for your entire life and not tell it apart from perfect play. No-limit hold'em, heads-up or six-player, isn't solved. Programs like Libratus and Pluribus beat top professionals by a statistically significant margin, but beating people isn't the same as having a provably optimal strategy. Six-player poker also lacks the clean guarantee of heads-up, where an equilibrium strategy can't lose on average.Can ChatGPT or another chatbot play poker well?Not at a solver level. In the PokerBench benchmark presented at AAAI 2025, GPT-4 was the best general-purpose model and still matched the optimal action and bet size in only 53.55% of 11,000 test spots. Other models did worse, and fine-tuning on poker data helped. A general language model doesn't run game-theory calculations on its own. It can explain concepts and walk through a hand in plain language, which is useful for learning, but dedicated poker AI systems are trained differently.Are AI poker bots stronger than rule-based bots?In the only public bot-versus-bot comparison of commercial products we know of, a 2026 preprint, yes. Over 17,295 hands in late 2025, two machine-learning bots won +29.2 bb/100 as a group, while four rule-based configurations lost −24.3 bb/100, and the confidence intervals didn't overlap. Even a paid premium profile left the rule-based approach at −28.0 bb/100. The bots played each other at micro stakes with no humans involved, so the results say nothing about play against people.Which poker bot is the best in 2026?There is no independent ranking of poker bots, but the available data points one way. The strongest approach in 2026 is a hybrid: a specially trained neural network combined with game-theory algorithms, the design used by products such as PokerBotAI. Older rule-based bots, which run on hand-written player profiles, look weak on the numbers. In the 2026 bot-versus-bot study, every rule-based configuration tested finished below zero over 17,295 hands, including one with a paid top-tier profile at −28.0 bb/100. In that test, a buyer of any of them would have paid a license or subscription fee for a program that lost over the long run. Before paying for any bot, ask the vendor for long-term results with confidence intervals.
Reactions
0
0 Comments
Be the first to comment
Nothing here yet.