How we trained the GTO MAX AI exploit model on 304,000 real hands

1. Why the old exploit line felt too careful

GTO MAX AI shows two lines on every decision: the GTO line, and an Exploit line built from a read on the player in front of you. Players kept telling us the same thing about the second one: it agreed with the GTO line far too often, even against opponents with obvious leaks.

They were right, and the reason was caution stacked on caution. Every read was shrunk toward a "solid regular" once for its sample size, then again for how sure the stat was, then again for the profile's overall confidence. A player who folded 70% of the 12 c-bets they faced earned about 11% more flop bluffs. Against that player the right adjustment is several times bigger.

We also found two counting errors. "Went to showdown" was divided by every hand dealt instead of the flops a player saw, so 97% of players looked like they give up before showdown, and the old model added turn and river bluffs against almost everyone, calling stations included. Fold to c-bet and fold to river bet only counted a player's first action on the street, so every check-then-fold out of position was missed.

2. The test: every read checked against the next hand

To fix this properly we needed ground truth, so we took 304,000 real online cash-game hands from a public research dataset of hand histories: no-limit hold'em at $0.10/$0.25, 6,462 players with 50 hands or more. We replayed them in order through the app's own profiling code. Before each hand, the model made its read on every player, using only the hands that came before. Then we checked that read against what the player actually did.

That gives a hard test of every prediction. If the model says a player folds to river bets 60% of the time, the players it says that about should fold close to 60% of the time in the hands that follow.

The old reads failed it. A player's own past fold rate predicted their next fold much less than the model assumed: it trusted six to eight chances where about 32 were needed before a player's own numbers deserve as much weight as the table's average. The reads were confident where they should have been modest, and modest where they should have been bold.

3. What actually predicts a fold

The replay answered a better question than "who is this player": what makes anyone fold? Three things stood out.

• The bet size. The pool folded 35% to a third-pot river bet, 50% to half pot and 73% to an overbet. A pure bluff with no showdown value earned +0.17 to +0.33 of the pot, more at bigger sizes. Size, with the table's own habits, explained more of a fold than any one player's stats.
• The line. A player who had not called a bet yet in the hand folded 45-69% to a turn bet; once they had called the flop, only 20-48%. A turn bluff earned about +0.35 of the pot in the first case and lost a little in the second. On the river, bluffs paid on every line but less each time the player had already called: about +0.44, +0.22 and +0.15 of the pot.
• The table. Players fold, bluff and raise like the pool they come from. In this pool, called river bets were bluffs only 21% of the time (14% for overbets) when a bluff-catcher needs 27-38% from half pot up; raises seen at showdown were value 75-89% of the time; and players folded to 3-bets 41% of the time, not the 55-60% the old model assumed.

A player's VPIP says something about their folding too: looser players fold a little less, even after their own fold numbers are counted.

4. What we rebuilt

The new exploit model is built around those findings. Each change was tested on the same real hands before it shipped.

• It learns your tables. The app now keeps a running read of the whole pool you play against: every player you have seen, how they respond to each bet size on each street, how often their river bets and raises turn out to be bluffs. A new player starts from that pool, not from a textbook regular, so the pool's leaks are exploited from the first hand.
• It trusts each player the right amount, once. A player's own numbers move the read as fast as the replay showed they should, and there is no extra damping on top.
• Every exploit is priced in pots. A bluff is worth what it wins: the folds it gets at that size, minus what it loses when called. The shift in the Exploit line grows with that value, and the bet goes in at the size that wins the most against that player.
• Bluffs read the line. More stabs when your opponent has not called a bet yet, fewer second barrels into a player who just called the flop.
• River and raise decisions read the table. Bluff-catchers are priced separately against small and big river bets, and medium hands fold more often to raises when the table's raises are value.
• Preflop reads apply where they belong. A steal read applies only when the hand is folded to you in late position, and a read on a player's opens applies only when you face their open.

The AI read that takes over after a couple of dozen hands now sees the pool and the priced read, and keeps its bet sizes.

5. How we tested it: 12,000 simulated hands

A replay can measure each read. It cannot tell you what the whole strategy wins, because the cards and opponents would play out differently once your decisions change. For that we built a full six-max simulation.

The hero is the app itself: the same preflop charts, the same flop engine and the same live solver on the turn and river that you use, with the same player profiling running in the background. Its five opponents are fitted to real players from the dataset: their preflop ranges, how often they fold to each bet size on each street, how often they bluff and raise, and how they play when the hand gets big. Information is imperfect both ways: the app sees only what you would see at the table, the opponents see only their own cards, and folded hands are never revealed. Every pot is raked at 5%, capped at 3 big blinds. All-in hands are scored by their equity rather than the run-out, to take the luck out.

The old and new exploit lines played exactly the same deals against exactly the same opponents, so the comparison is paired hand by hand:

The simulation also caught a mistake the replay could not see: the old steal read applied from any seat, so a tight player behind you made the app open wider under the gun. That is fixed.

6. What you will see at the table

The GTO line is unchanged. The Exploit line beside it now moves much further when there is a real edge, and it tells you why in the reason under it: what the pool or the player folds at that size, what a bluff needs, and what it earns.

• Bluffs into players and pools that over-fold, often at a bigger size than the GTO line uses.
• Fewer bluffs into calling stations once their own numbers prove it.
• Turn stabs when your opponent has not called a bet yet in the hand, and more turn checks after they called your flop bet.
• More folds with bluff-catchers against big river bets and raises at tables where those are rarely bluffs.
• Value bets that bet more often against players who pay.

7. The limits

Simulated opponents do not adapt, and real players sometimes do: if a regular notices you bluffing every river, the read will follow their folds, but it will take some hands. The replay data is micro-stakes cash from an earlier era of online poker, not your table, which is exactly why the model now learns the pool you actually play against instead of carrying fixed numbers. And no model is right every hand; the reason under each line is there so you can check it.

GTO MAX AI is a coaching and study tool. Use it where real-time assistance is allowed: study, home games, private clubs and play money. Every decision is yours. If you want the background on when to leave the equilibrium at all, read GTO vs exploitative poker, and for how profiles are built, see poker player profiling.

More from the blog

• New preflop ranges: what changed in the RTA software this week, with the numbers
• AAPoker is now fully supported: a real-time GTO coach for AAPoker tables, in the browser
• CoinPoker is now fully supported: a real-time GTO coach for CoinPoker tables
• AI vision capture for poker: why GTO MAX AI reads the table faster and more accurately than any RTA

Longer reads: the poker RTA guide and all guides.