BLOG

AI in poker: what RTA, bots and solvers actually do, and what they do not

GTO MAX AI team · 4 October 2026 · 7 min read

Every poker tool now calls itself AI. Some are, in the sense that matters; most are a solver with a new coat of paint; a few are a chat model guessing. These are different machines with different failure modes, and a player choosing between them should know which is which. This is a plain account of what each one computes.

1. Solvers: the equilibrium, offline

A solver (PioSOLVER, GTO+, TexasSolver, the engines behind GTO Wizard) takes a precisely defined game, two players, fixed ranges, a fixed bet-size tree, and iterates counterfactual regret minimisation until neither player can gain by changing strategy. The output is a strategy for every hand at every node: frequencies, not instructions. It is deterministic, exact to a stated tolerance, and slow: a flop subtree takes minutes and gigabytes. Nothing about it is "intelligent" in the everyday sense. It is a very large calculation, and its answers are only as good as the ranges and sizes you gave it.

2. Bots: a solver that clicks

A bot reads the table and acts, with no human in the loop. The reading is computer vision; the deciding is a solver library or a trained policy; the acting is automation. Bots are what every site's security team is built around, because they scale: one operator, forty tables. They are also detectable by the same statistical fingerprint as any solver follower, plus timing, plus the inhuman uniformity of the clicking. Every reputable tool in the RTA space draws the line here: an RTA advises, the human acts. The moment software clicks, it is a bot, legally and practically.

3. Real-time assistance: the solver, during the hand

An RTA brings solver output to a live decision. The hard part is not the solving, it is the reading: turning a screen into a hand state (cards, board, stacks, bets, pot, who has folded, whose turn) in under a second, on a client that animates and moves things. Most RTAs are a pixel reader plus a precomputed library of solutions; the better ones solve turns and rivers live with the actual ranges. The output is still the equilibrium. Where that sits on the AI spectrum: the reader uses computer vision models and the solver uses game theory, and neither has anything to say about the person across the table. How this works step by step: how a live solve works.

4. Opponent modelling: the part that is actually about the opponent

The equilibrium is the same against everyone, which is both its strength and its limit. Opponent modelling estimates how a specific player deviates from it, from the hands observed, and adjusts. The honest way to do this is statistical: each tendency as a posterior that starts at a population prior and updates with every observed decision, with the uncertainty kept, so three hands buy a small lean and two hundred buy a conviction. This is where the useful "AI" in a poker tool lives, because it is the only part that produces something a solver cannot. It is also where the marketing is loosest: "adaptive" and "learns your opponents" are claims to test with a question: how many hands before it adjusts, and by how much?

5. Large language models: explanation, not computation

GPT-5-class models are now in several poker products, and it matters what they are used for. A language model does not compute equity or equilibrium; ask one to solve a river and it will produce a confident paragraph that is wrong in the numbers. What it is good at is turning numbers into an explanation: given this player's statistics, this spot's equilibrium and your notes, write the reason for the adjustment in words. That is how GTO MAX AI uses one. The arithmetic is done by the solver and the profile; the model writes the sentence. A tool that lets a language model make the decision is a tool that has not understood what language models are.

6. Vision models: reading the table

Multimodal models can read a screenshot of a table, and early RTAs used them for everything. They are slow (seconds), expensive (per frame) and occasionally wrong in ways a pixel rule never is (a 6 read as a 9). The current approach is to read everything on device, cards from learned templates and numbers from OCR, and ask a vision model only for a glyph nobody has seen before, once, then learn it. Latency falls to milliseconds and the cost to nearly nothing. If a product says "AI vision" for every read, ask how long a read takes.

7. How to tell what you are buying

"Solved with AI": it is a solver. Ask which engine, which tree, which accuracy.
"AI reads the table": ask whether every frame goes to a model, and how long a read takes.
"Adapts to opponents": ask how many hands before it adjusts, what it tracks, and whether you can see the numbers.
"AI explains the play": fine, as long as the numbers underneath came from a solver.
"AI plays for you": a bot. Walk away.

What the complete field looks like, tool by tool: the poker RTA guide.

GTO MAX AI: the RTA that knows the player.

GTO line, exploit line and the reason, live, on up to four tables, in Chrome. GGPoker, Natural8, ClubGG, PokerNow, and PokerStars in beta. Try it for 30 minutes on Telegram.

Get a 30-minute trial

Coaching and study tool. Use it where real-time assistance is allowed: study, home games, private clubs, play money. Most online sites forbid real-time assistance. Advice can be wrong; every decision is yours. Terms of use.

More from the blog

Longer reads: the poker RTA guide and all guides.