Poker Is Solved. A Startup Just Shipped the Proof.
Chess was cracked in 1997, and it took the world two years to download the engine. Poker's lab proof came in 2019 — and sat in a university for seven years. Now a startup has shipped the engine. Here's what it does on real work, and where it stops.
The pattern from chess
In 1997, Deep Blue beat Kasparov, and the world did what it always does at a lab result: called it impressive, decided it was a machine and not a tool, and went back to playing. Then, a few years later, grandmaster-grade analysis became a download — then a free app on every club player's phone. And the game changed shape: every level strengthened at once, the casuals got squeezed, and the players who trained with the engine pulled away permanently. The late adopters never caught up.
The gap between "impressive lab demo" and "an engine in your hands" took about two years in chess. The people who were inside that gap collected the money.
Poker just ran the first half of that pattern. A startup just closed the second.
The proof — and the seven-year gap
Poker was supposed to be immune to the machines. Chess is a fixed board with perfect information, ~10⁴⁷ states. No-Limit Hold'em has two hidden cards, bluffing as a core mechanic, and a game tree of roughly 10¹⁶⁰ reachable states. For decades the answer to "can a machine play poker?" was "the game is different."
It wasn't different enough. The proof came in three lab breakthroughs:
- 2015 — UMass mathematicians solve Limit Hold'em. A rigorous proof that poker is a solvable game.
- 2017 — Libratus (Carnegie Mellon) beats a world champion heads-up No-Limit.
- 2019 — Pluribus (Facebook) plays six-max against four of the world's best human players — four on one — and wins the majority of the hands. The exact table shape you play every night.
And then the gap. For seven years, the engine that beat four pros was research — a paper, a server rack in a university, a feeling in the community that the people who get this engine will be untouchable. The field's toolkit stayed exactly where it was: charts (snapshots of someone else's assumptions), pre-solved libraries (their spots at their sizings and their rake), and desktop solvers — PioSolver, MonkerSolver — at the cost of a $249–$1,100 license, 64–128GB of RAM, and hours of convergence per tree. A part-time job before the first answer.
"People who get this engine" meant people with a lab.
What the startup shipped
MonsterSolver is the startup that just closed the gap — a state-of-the-art GTO poker software: a neural solving engine running on cloud GPUs. You give it a spot — exact hand, exact board, exact stacks, any action history, 2 to 6 players — and it returns the GTO-optimal strategy for that precise situation in seconds: a complete range grid with action frequencies and EV for every line.
What changed is the business model as much as the math:
- Nothing to run. No tree building, no Java flags, no 64–128GB of RAM. The compute lives in a data center, not your basement.
- No license fee. A small prepaid balance — pay only for what you solve, free credits when you sign up.
- No plumbing. It reads the table itself: this RTA poker software works with virtually every client, no per-room plugins, no hand-history files. It sees what you see, on whatever site you sit at.
- Any spot. Preflop and postflop, any board, any sizing, 2 to 6 players — the HU flop through the six-handed monster.
Their about page says it best: "fast, affordable and available to everyone — not just pros with a server rack in the basement." One line, and it's the whole story: the lab's monopoly on the engine just broke, and it broke toward a billing page.
Tested on real work: three spots you can run yourself
The engine's claims are checkable, which is how we test any tool. Three tests, in order of effort:
Test 1 — the exact spot. Take the last hand where you weren't sure what you were doing. Type it in: your cards, the board, the stacks, the action so far. Read the grid, not just the recommended action: which hands bet, which check, which fold, and what each line is worth. If the engine disagrees with what you played, the gap is priced in bb. That's your leak, with a number on it.
Test 2 — the session. Point it at the session you just played. Every spot gets solved against your exact ranges, your exact stacks, your exact action lines — before you've made coffee. A 300-hand session, graded, in the time the kettle takes.
Test 3 — the multiway. Take a three-way flop you face every week. The heads-up solvers in most stacks can't see that table at all — 2 to 6 players is this engine's home ground, the same topology Pluribus proved in 2019.
What people consistently find: bet sizing turns out less "style" than they thought, one c-bet range is off in a predictable direction, and check-raise frequencies on dry boards don't match the number they carried for years. The pattern matters more than any single hand — and that's what compounds.
"Solved" — the honest version
The number is big enough to hide imprecision, so here's the precise claim: nobody printed the complete strategy for No-Limit Hold'em — 10¹⁶⁰ states means the full solution isn't coming. But chess was never "solved" either — it was cracked. The day amateur-runnable analysis arrived, the best human in the world quietly became the second-best player in the room. Nobody needed the complete solution; they needed the answer key for the position in front of them.
That's what this startup ships. And the approximations are already strong enough that no human plays closer to equilibrium than the software — Pluribus beat four pros on exactly that strength. Every solve is a slice of it, pointed at the spot you're holding.
Where it stops (say it out loud)
Full disclosure: MonsterSolver is RTA-capable. It reads the table on virtually every popular poker site and returns an answer in 1–5 seconds to each situation — live, in real-money games.
We don't recommend using it that way. It's not fair — no opponent at your table has an answer key, and MonsterSolver's own fair-use policy says study and training only, never real-time assistance in real-money play. And it's non-human — that's the tell. You might never get caught, but use it often and you'll always win, and a consistent, humanly-impossible win rate is exactly what detection teams look for. When they do flag you, the penalty at every legitimate room is the same: permanent closure, fund seizure, blacklist. Don't be the case study.
The three uses that are worth the money:
1. Training — drill spots in play-money, instinct first, solve second, until the answer stops surprising you. 2. Post-game analysis — the session you just played, every spot solved against your exact ranges, before the coffee's done. 3. Streaming — run the solve on stream and let your chat watch the best move appear, the way televised poker commentary calls the "right" play. The audience gets the analysis; the table stays honest. (And content creators take a 30% referral share — the distribution of the tool is its own business model.)
The software doesn't beat poker geniuses at their table. It beats them in the training room — every single session.
The math of a startup
The old entry cost for real solver study was a stack: a $249–$1,100 license, a machine upgrade, a learning curve in weeks. The startup's entry cost is a small balance of prepaid credits — pay only for what you solve, free credits at signup. One real leak found in week one pays for the lifetime.
The two years
Chess ran its transition in two years: 1997 in the lab, 2008 on every laptop. Poker's transition had a seven-year head start in the lab — and a startup just collapsed the second half to zero. The transition is running now, and chess history says two things about it: the pool is softest during the wave, and the players who train with the engine pull away permanently.
The engine that beat four pros in a lab is now a billing page. The table stays human.
The only question left is which side of the board you're on.
