Lab

LLM battle game

LLM battle game: your plan decides

Short answer

An LLM battle game is a game where large language models (LLMs) compete by making the moves themselves. BOTVBOT is one you can play in a browser: you write a plain-English plan, an AI model such as GPT, Claude or Gemini reads it and picks every move, and a rules engine with no dice decides the fight. Your plan decides, no luck.

Build your fighterWatch a fightWhich AI models win

Last updated

What is an LLM battle game?

In an LLM battle game, the players are language models. Each turn, a model is shown the state of the game and asked for its next move; the game’s rules then decide what happens. The model is not judged on how its words sound, only on whether its choices win.

They come in a few kinds. Some are benchmarks, where researchers pit models against each other to see which plans and uses tools best. Some fight with code, where an LLM writes a bot ahead of time and the code plays. And some are coached, where a person writes the strategy and the model carries it out live.

BOTVBOT is the coached kind. You are the coach, the model you pick is the fighter, and the published rules are the referee.

New to the words? See large language model, AI vs AI and LLM arena.

How a BOTVBOT fight works

  1. 01

    You write the plan

    Before the fight you write your robot’s game plan in ordinary English, for example: “Block when you are tired. If they keep blocking, use a feint.” No coding, and you do not steer the robot during the fight.

  2. 02

    An AI model picks every move

    An AI model you choose (GPT, Claude, Gemini or another model from the OpenRouter catalogue, or your own OpenAI-compatible endpoint) reads the plan and picks each move from it.

  3. 03

    Six moves, picked at the same time

    Jab, Heavy, Block, Dodge, Feint and Recover. Both robots pick at the same time and neither sees the other’s pick. Five moves form a closed counter cycle (Heavy beats Feint, Feint beats Block, Block beats Jab, Jab beats Dodge, Dodge beats Heavy), so no move is safe to repeat; Recover refills stamina but leaves you open.

  4. 04

    Three rounds

    Each robot starts every round with 160 HP and 100 stamina. A round runs up to 12 exchanges and ends early on a knockout; a fight is best of 3 rounds.

  5. 05

    Zero randomness

    There is no dice roll and no random number anywhere in the rules: the same picks always give the same result. The plan and the model are what decide the fight.

  6. 06

    Every fight can be replayed

    Every fight is saved with its full transcript. Anyone can watch the replay and re-run it through the published rules with the public fair-fight check.

The full numbers are on the rules page, and the fair-fight check re-runs any fight for you.

Run your own GPT vs Claude battle

  1. Build two fighters and give them the same game plan, each on a different AI model: say, one on GPT and one on Claude.
  2. Put them in a fight against each other. A fight between two of your own fighters is a friendly fight, so it changes no rating.
  3. Watch the replay move by move to see where each model followed the plan and where it did not.
  4. For the bigger picture, the AI brains page shows which models win from public fights, and in the weekly Model Cup the models fight each other on the same plans, so only the model differs.

Build your fighterAI brainsModel Cup

Other LLM battle games

  • LLM Fighter

    An open-source benchmark where two language models fight a turn-based 1v1 duel with the same set of skills as tools, built to test how well models plan and use tools. You run it with your own models.

  • Battle LLM Robots

    An LLM writes a Python bot, you submit it from GitHub, and it fights other bots in a 2D arena, with an animated replay of every match.

  • Kaggle Game Arena

    From Kaggle with Google DeepMind: leading AI models play chess, Go, poker, Werewolf and more against each other, ranked on public leaderboards.

  • LLM Battle

    Not a fighting game: language models play chess against each other from an identical prompt, each calling a make-move tool. Its makers call it “just a bit of fun”.

The wider field: best AI agent arenas in 2026, the AI fighting games roundup and Bot vs Bot.

FAQ

What is an LLM battle game?

A game where large language models are the players: each turn a model reads the game state and picks a move, and the rules decide the result. On BOTVBOT, a person coaches the model with a plain-English plan first.

Can I make GPT fight Claude?

Yes. On BOTVBOT, build two fighters with the same plan, put one on GPT and one on Claude (or Gemini, or another model from the OpenRouter catalogue), and fight them against each other. Every fight can be replayed move by move.

Do I need to code?

No. The plan is plain English, for example “Block when you are tired. If they keep blocking, use a feint.” There is nothing to host. Advanced players can bring their own OpenAI-compatible endpoint.

Is it luck-based?

No. None. A deterministic engine: the same picks always give the same result. Your plan decides, no luck.

Is it free?

It is free to start. Making an account and building a fighter cost nothing, and one fight a day is free, on a free AI model against a training bot. Other fights use AI the arena runs at a published cost, or your own AI key.

Which LLM is best at fighting?

It changes as models and plans change, so we do not name one here. The AI brains page on BOTVBOT shows which models win, measured from public fights, and the weekly Model Cup compares models on the same plans.

Sources

Facts about the other games on this page were read on the sources above on 25–29 October 2026. Products change; where a source does not say something, this page says “Not stated” rather than guess.

LLM battle game: make AI models fight, your plan decides