GTO MAX AI is now trained on 5 million real hands

1. What the 5 million hands are

GTO MAX AI is now trained on 5 million real hands of online six-max no-limit hold'em, from NL5 to NL100, the stakes most of our players play.

Every hand is kept as a full record: the seats and their stacks in big blinds, the positions, every action with its size on every street, the board, the cards shown at showdown, and who won what after rake. Every hand that reached a showdown has every player's cards, the losers' included. That matters more than it sounds: it means we can see what players actually hold when they bet, call and raise, not just what they did.

Earlier this month we rebuilt the exploit model on 304,000 hands from a research dataset. The full 5 million hands cover five stakes instead of one, and most of them were played in the last two years.

2. A table of real players at every stake

The first thing the hands built is a simulator. For each stake from NL5 to NL100, we fitted opponents to how the real players at that stake play: how often they open, call and 3-bet from each seat, how often they c-bet, how they react to each bet size on each street, how often they bluff and how often they give up. Then GTO MAX AI sits down at those tables and plays, exactly as it plays for a customer: the same preflop charts, the same solved flops, the same live turn and river solves, the same exploit line.

The important part is how two versions of the app are compared. Both play the same deals, with the same cards and the same opponent choices wherever their decisions agree, so when one version wins more it is because of what it did differently, not because it was dealt better cards. A proposed change becomes a what-if version of the app, and it plays 50,000 deals, 10,000 at each stake, against the current version on those same deals before anyone decides to ship it.

3. What it changed: strong hands on the turn

The first change the simulator tested came from our own players: the exploit line sometimes folded strong hands on the turn against opponents it read as rarely bluffing. Against real raise ranges those folds cost money. With the turn's folding shifts switched off, the app won more at every stake:

That change shipped in version 7.04. Not every idea passes. Betting half of the heads-up turns the GTO line checks, at three quarters of the pot, lost 2.3 bb/100 in the same test, so it stayed out of the app. A quarter of them came out at +0.2 bb/100, too close to call, so it stayed out too.

4. What it changed: facing a bet and a raise

The second change is about the most expensive spot in multiway pots: one player bets, another raises, and you are next. The app used to read that raise with a hand-set table that kept draws almost as often as made hands. On wet multiway turns, that table assumed a raise was about a third two pair or better and 44% draws.

The hands say otherwise. We took 29,240 raises from the dataset that went to showdown and fitted how often each kind of hand raises, given what the board makes possible, heads-up and multiway, on every street and for small and big raises. On wet multiway turns, raisers showed two pair or better 61% of the time and draws 20%. The fitted model says 59% and 21%. Tested on hands it had not seen, it predicted what raisers hold better than the old table on every street, heads-up and multiway.

A customer's hand shows the difference. Top pair, king kicker, facing a bet and a raise on 9-8-7-4 with two clubs, was told to call every time with "47% equity". Against the ranges that really raise there, it has about 25%, and the price needs 35%. Version 7.06 folds it, and decides every raise with two or more opponents in on the price, with a margin for each player still to act behind. Sets, two pair and the nuts still continue, and the nuts raise.

5. Machine learning on the exploit line

A whole hand is a noisy way to judge a single decision, so the simulator can also probe one. At an exploit decision, every option the line lists is played to the end of the hand many times, on the same rollouts of the cards to come and the opponents' choices. That gives each exploit shift a measured gain over the GTO mix at that exact spot, with far less noise than comparing whole hands.

From those measurements, a gradient-boosted model learns when a shift earns and when it does not: by street, position, the kind of spot and how many hands of evidence stand behind the read. Its verdicts are scored on decisions it did not train on, beside simple rules such as "only shift after 30 or 60 hands on the player", so we can see whether the learned gate beats the plain one before it touches the advice.

6. What 5 million hands cannot do

The simulated opponents are models of real players, fitted to how they act on average, not the players themselves. They do not tilt or adjust to you, and a stake's pool on one site is not the same as another site's. So we use the simulator to compare versions of the app against each other on the same deals, where those effects cancel out. We do not use it to promise a win rate.

The dataset also does not replace the read on the player in front of you. The exploit line still starts every opponent at the pool's average and moves toward their own numbers only as their hands come in, as the player profiling guide explains. And no dataset turns a multiway pot into a solved game: no solver covers three or more players, so those spots stay on the range-and-equity engine, now with raise ranges measured on real hands.

7. What you will see at the table

• Strong hands kept on the turn. Sets and overpairs facing turn aggression continue where they used to fold against "tight" reads.
• Fewer cold calls of a bet and a raise. One pair facing a bet and a raise with three in folds unless the price is there, and the note says what equity the call needs.
• Equity against real raise ranges. The equity shown against a raise is computed against the hands raisers actually show, on that board.
• Changes that earn first. Each change to the advice is measured on the same deals against the version you use today before it ships.

The GTO line itself is unchanged: preflop from the solved charts, flops from the solution library, turns and rivers solved live. What the 5 million hands changed is how the app reads the players and how it decides when to leave the equilibrium.

More from the blog

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

Longer reads: the poker RTA guide and all guides.