Real-time poker solver: how a live solve works, street by street

1. Why you cannot solve a whole hand at the table

A no-limit hold'em hand from the flop has a game tree with tens of millions of decision nodes once you allow a few bet sizes per street. Solving it to a useful accuracy takes a counterfactual-regret solver minutes on a strong server and several gigabytes of memory. Your clock gives you 15 to 30 seconds. So no tool solves a flop while you wait, whatever the product page says. What they do instead is the whole design question.

2. What happens on each street

Preflop: charts

Preflop is a solved, finite problem per format. GTO MAX AI uses position-by-position charts for opens, 3-bets, 4-bets and calls, by stack depth, with push/fold charts taking over under 15 big blinds.

Flop: a library, filled by use

Flops are solved ahead of time and stored. A stored spot is identified by the flop in canonical suit form (there are 1,755 distinct flops), both players' ranges as the flop begins, which follow from the preflop action, and the stack-to-pot ratio in buckets. The standard single-raised pots (UTG, HJ, CO, BTN and SB opening against the big blind, and CO or HJ against the button) are pre-loaded. A flop that is not in the library yet is answered instantly by the rules engine and queued on the server, which solves it in the background, a few minutes per spot, and stores it permanently. The next time that flop and situation comes up, for any user, the answer is the solved one.

Turn and river: solved live

From the turn on the tree is small enough. The moment the turn card is on the table, the whole street is queued on the solver pool with both ranges as they stand, so by the time the action reaches you the solution is normally there and your node is found by walking the street's action through the tree. A river solve takes a fraction of a second, a turn a few seconds. While a solve is still running the rules-based line stands in, marked as pending, and is replaced when the solution lands.

3. The tree: sizes and accuracy

Every solve is only as good as its tree. GTO MAX AI's turn tree allows bets of 33% and 75% of the pot on the turn and 75% on the river, raises at 75% of the pot, and all-in at every node. The river tree allows 33%, 75% and 125%. A real-world size that is not in the tree is mapped to the nearest size that is, which is standard practice and a known limitation: a 150% turn overbet is read as a 75% bet for solving purposes.

Accuracy is measured as exploitability: how much an opponent who knew the strategy exactly could win against it, as a share of the pot. The turn solves stop at 2% of the pot, river solves at 0.8%. For comparison, study solvers are usually run to 0.3 to 1% and nobody at a table plays within 5% of equilibrium.

The tool's own hand is always in its range. A solver that drops your actual holding from the range because the equilibrium would have folded it earlier gives you no answer for the hand you are holding; GTO MAX AI keeps a floor weight on it.

4. Multiway pots: where the solver stops

A CFR solver works on a two-player tree. Three or more players postflop multiplies the tree beyond anything that can be solved live, and no RTA on the market solves it. In those spots GTO MAX AI uses a rules engine instead: it computes your equity against each remaining player's actual range, checks range advantage, board texture, position and stack-to-pot ratio, and picks a line with the adjustments multiway theory prescribes (more checking, smaller bets, fewer bluffs). It is grounded in the same inputs a solver uses, but it is an approximation. Expect it to be right on the clear decisions and less precise on the thin ones.

5. What sits on top of the solve

Equilibrium is the floor. GTO MAX AI's second line, the exploit line, moves off it in the direction the opponent's profile justifies, sized by how far that player sits from indifference and damped by how many hands the model has on them. How that model is built is its own guide: player profiling and the exploit line.

6. FAQ

Can a poker solver really run in real time?

Turns and rivers can be solved live, in seconds, with a small enough tree. Flops cannot: they take minutes and gigabytes, so every RTA uses a precomputed flop library. Any tool claiming to solve flops live while you wait is describing a lookup.

How accurate is a real-time solve?

GTO MAX AI stops turn solves at 2% of the pot exploitability and river solves at 0.8%. Study solvers are usually run to 0.3 to 1%. The difference is far below the error of any human strategy.

What bet sizes does the live solver use?

Turn: 33% and 75% of pot, river 75%, raises 75%, all-in always. River: 33%, 75% and 125%. A size outside the tree is mapped to the nearest one in it.

What happens in a multiway pot?

A CFR solver is two-player. Multiway pots are handled by a rules engine that uses equity against each player's range, range advantage, texture, position and stack-to-pot ratio. It is an approximation, strongest on clear decisions.

Do I wait for a flop that is not in the library?

No. You get an instant answer from the rules engine, the server solves the flop in the background in a few minutes, and the solved answer is stored for every future occurrence, for every user.

Related guides

• Poker RTA explained: What real-time assistance is, manual vs automated, how a table is read and solved, where it is allowed, how sites detect it, the tools compared.
• Poker RTA software: A buyer's guide: how the software works, six things to test in a trial, what it costs, how the tools compare.
• Player profiling: How the exploit line is built: a Bayesian posterior per tendency, archetypes, the 20-hand rule, your own notes.