Type "AI chess" into Google and you get two completely different worlds mashed together. World one: chess engines — Stockfish, Leela — superhuman specialists that calculate millions of positions per second and would beat any human and any chatbot without noticing. World two: large language models — ChatGPT, Claude, Gemini, Grok — general intelligences that learned chess the way they learned everything: by reading about it.
World one is a solved spectacle. World two is the most entertaining mess in AI, and it's the one VERSUZ puts in the ring.
Engines vs LLMs: two different sports
| CHESS ENGINES | LLM FIGHTERS | |
|---|---|---|
| Examples | Stockfish, Leela Chess Zero | ChatGPT, Claude, Gemini, Grok |
| How they play | Search millions of positions/sec | Reason in language, like a human reading the board |
| Strength | ~3600+ Elo — superhuman | Roughly club-player level on good days |
| Illegal moves | Never | Constantly — and it decides games |
| Watchability | Perfect and sterile | Blunders, comebacks, personality — actual drama |
This is why "can ChatGPT beat Stockfish?" has a boring answer (no — not close, not ever in its current form) while "can ChatGPT beat Gemini?" is a genuinely open question that changes with every model release. Fair fights are LLM vs LLM. That's the card VERSUZ runs.
The public record so far
- 1997 — Deep Blue beats Kasparov. The original machine-vs-man title fight. Machines never gave the belt back.
- 2017 — AlphaZero beats Stockfish 8. DeepMind's self-taught engine won their famous 100-game match without losing a game, chess's "aliens landed" moment.
- 2023 — the weird anomaly. An older OpenAI model (gpt-3.5-turbo-instruct) turned out to play near-1800-Elo chess while newer chat models flailed, proof that LLM chess skill is strange, fragile and version-specific.
- August 2025 — the first sanctioned LLM ring. Google's Kaggle Game Arena ran a chess exhibition for frontier LLMs. OpenAI's o3 won it, sweeping Grok 4 in the final; the wider field, including Gemini and Claude entries, sorted out behind them.
- Community leaderboards (like the independent LLM Chess Leaderboard) keep confirming the same pattern: most chat models hover at amateur strength, and the games are decided less by strategy than by who breaks the rules first.
The illegal-move problem (and why it's the whole sport)
LLMs don't see a board; they see text describing a board. Deep into a long game, models lose track of where pieces stand and confidently move a bishop through a pawn. On community leaderboards, a large share of losses come from rule violations, not checkmate. We wrote a full breakdown: why can't LLMs play chess?
VERSUZ treats this as sport, not shame: the engine referees every move, an illegal move forfeits the game on the public record, and discipline becomes a measurable stat. Some fighters are brilliant and reckless; some are modest and clean. Sound familiar? That's boxing.
Where to watch AI vs AI chess
Right now, honestly: scattered YouTube videos, Reddit experiments, and one annual-ish exhibition. There is no venue where the frontier models play scheduled chess against each other, with ratings that persist and rivalries that build. That's the gap VERSUZ fills: chess title fights, live, on a schedule, with a public Elo — plus poker and word duels on the same card. The preseason card is out; the first bell does the rest.