GTO stands for “game theory optimal.” In poker it means playing a strategy close to a Nash equilibrium: a way of playing that an opponent cannot exploit, because no change in their strategy would make them more money against it. A GTO player is not trying to read anyone. They are trying to be impossible to read for profit.
That is the whole idea. The rest of this note explains where it comes from, what the math looks like on a single river decision, what the landmark poker AIs actually proved, and why most players at a $1/$2 table should treat GTO as a map, not a script.
Nash equilibrium, in one paragraph
The term comes from John Nash’s work in game theory. Noam Brown and Tuomas Sandholm, the Carnegie Mellon researchers behind the Pluribus poker AI, define it neatly in their 2019 paper in Science: a Nash equilibrium is a list of strategies, one for each player, in which no player can improve by deviating to a different strategy.
Their example is rock-paper-scissors. The equilibrium is to pick each option one third of the time at random. Against that strategy, the best any opponent can do in expectation is tie. It cannot be beaten. It also does not win anything extra against a player who throws rock every time; that is the trade-off at the heart of GTO.
Why poker is the perfect game for it
Poker is a game of hidden information. Brown and Sandholm note that the foundational papers on game theory used poker to illustrate their ideas, because no other popular game captures hidden information as cleanly. A player who always bets strong hands and always checks weak ones is as readable as the rock thrower. The fix is to mix: bet some strong hands, some bluffs, and check some of both, in proportions that leave the opponent with no profitable adjustment.
The math on one river bet
Here is where “balanced” stops being a buzzword. This is a toy model, the standard textbook simplification, not output from any solver.
The river is dealt. The hero’s range is now either the nuts or air (a busted draw). The opponent holds a bluff-catcher: a hand that beats every bluff and loses to every value hand. The hero bets b, measured as a fraction of the pot.
How often should the hero bluff? Just often enough that the opponent’s call breaks even. The caller risks b to win 1 + b (the pot plus the bet), so calling breaks even when the bettor is bluffing b ÷ (1 + 2b) of the time.
How often should the opponent call? Just often enough that the hero’s bluffs break even. A bluff risks b to win 1, so it breaks even when the opponent folds b ÷ (1 + b) of the time. The opponent must therefore continue 1 ÷ (1 + b) of the time. That figure is the minimum defense frequency.
| Bet size | Bluffs in the betting range | Value bets | Caller must defend |
|---|---|---|---|
| 1/3 pot | 20.0% | 80.0% | 75.0% |
| 1/2 pot | 25.0% | 75.0% | 66.7% |
| 2/3 pot | 28.6% | 71.4% | 60.0% |
| Pot | 33.3% | 66.7% | 50.0% |
| 2x pot | 40.0% | 60.0% | 33.3% |
Check the pot-sized row. The hero bets $100 into $100. The caller risks $100 to win $200, so a call breaks even if the hero is bluffing one time in three: 100 ÷ 300 = 33.3%. Two value bets for every bluff. The hero’s bluff risks $100 to win $100, so it breaks even if the opponent folds half the time, which means the opponent must call at least half the time to stop automatic profit.
Two lessons fall out of this little table:
- Bigger bets allow more bluffs. A 2x pot overbet can be 40% bluffs and still be balanced.
- Bigger bets require less defense. Facing that overbet, the caller only needs to continue a third of the time.
At this equilibrium, neither player can gain by changing frequencies. If the hero bluffs more, the opponent calls always and profits. If the opponent folds more, the hero bluffs always and profits. That knife edge is what “unexploitable” means.
What the poker AIs proved
Two papers in Science turned GTO from a theory into a working result.
Libratus (2017, heads-up). Brown and Sandholm’s “Superhuman AI for heads-up no-limit poker: Libratus beats top professionals” (Science 359, 2018) describes a 20-day, 120,000-hand match in January 2017 against four specialist professionals. Libratus won by 147 milli-big-blinds per game (thousandths of a big blind per hand) and beat each of the four individually. It worked by solving a simplified “abstraction” of the game for a blueprint strategy, then solving finer sub-games in real time as hands unfolded.
Pluribus (2019, six-handed). “Superhuman AI for multiplayer poker” (Science 365, 2019) moved to six-player no-limit Hold’em, the format most people play. Pluribus learned entirely by playing against copies of itself. Over 10,000 hands against elite professionals, it won an average of 48 mbb/game, with a standard error of 25. The blueprint strategy took 8 days on a 64-core server, 12,400 CPU core hours, which the authors put at about $144 at cloud spot prices then.
Two findings from Pluribus get quoted at poker tables everywhere:
- It confirmed the old advice that limping (just calling the big blind preflop) is a poor play for every seat except the small blind. It tried limping during training and dropped it.
- It disagreed with the folk wisdom that donk betting is a mistake, and did it far more often than professionals.
There is an important footnote, straight from the Pluribus paper. In two-player zero-sum games, an equilibrium strategy is guaranteed not to lose in expectation, which is why heads-up poker is the cleanest case. With three or more players, that guarantee disappears. The authors say their goal for six-player poker was not a specific game-theoretic solution concept, and that their algorithms are not guaranteed to converge to a Nash equilibrium outside two-player zero-sum games. Pluribus played a strategy strong enough to beat elite humans, which is a different claim. “GTO” in a six-handed game is, strictly, a slight overstatement of what anyone has.
What “GTO” means in a solver
When players say a solver “says” something, they mean software computed an approximate equilibrium for one spot under inputs the user chose: starting ranges for both players, stack depth, and a menu of allowed bet sizes. Change the inputs and the answer changes. A solver run is a careful answer to a precise question, and it is only as relevant as the question.
That is why careful study notes always list the settings. A strategy computed for 100-big-blind stacks with two bet sizes per street is not automatically right at a live table where everyone has 300 big blinds and the bet sizes are whatever someone grabbed.
GTO versus exploitative play
| GTO-based | Exploitative | |
|---|---|---|
| Goal | Cannot be beaten in the long run | Maximize profit against this opponent |
| Needs reads? | No | Yes |
| Risk | Leaves money on the table against weak players | Can itself be exploited if the read is wrong |
| Best against | Strong, unknown or adaptive opponents | Players with clear, repeated leaks |
The realistic approach for most students is to learn the equilibrium shape as a baseline, then deviate on purpose. If a regular in the game never bluffs the river, the hero should fold more bluff-catchers than the 50% defense figure says. The baseline shows how far that deviation is from neutral, and what it risks if the read is wrong.
A practical way to start
- Learn the defense and bluffing table above. It applies to every river.
- Start counting combinations; the ranges in poker note shows how.
- Pick one common spot (for example, button opens, big blind calls, flop checked to the button) and study it until the patterns stick.
- Only then add a study tool. The poker solvers comparison lists the desktop and cloud options by published price and format, and the GTO Wizard alternatives list covers free tiers.
One firm rule: study tools belong in study sessions. Real-time assistance during play breaks the rules of many poker rooms, so a player should read a room’s terms before running any software beside a table. The editors leave real-time assistance tools off recommendation lists entirely.
Sources
- Noam Brown and Tuomas Sandholm, “Superhuman AI for multiplayer poker,” Science 365 (6456), 885 to 890, 2019, DOI 10.1126/science.aay2400. Author’s copy, checked Nov 6, 2026.
- Noam Brown and Tuomas Sandholm, “Superhuman AI for heads-up no-limit poker: Libratus beats top professionals,” Science 359, 418 to 424, 2018. Author’s copy, checked Nov 6, 2026.
- River-bet frequencies: computed for this note from the break-even conditions shown above.
Poker is a game for adults, 21+ or 18+ depending on state law. Anyone worried about their play can call or text 1-800-MY-RESET for free, confidential help, and the responsible gambling page lists more.