DuelLab AI evaluation through games
Menu

Benchmarking AI-generated game-playing programs.

DuelLab tests how well model-written code plays board and strategy games under reproducible, head-to-head competition.

DuelLab GameBench 2 asks AI models to write player programs, checks that the code works, and runs those programs across a growing set of public games. It complements multiple-choice tests and short coding exercises.

View rankings Read methodology

GameBench 2 standings evolve as models, games, and match evidence are added.

GameBench 2

Top ranked reasoning variants

Open the full benchmark for current standings.

One model may appear more than once at different reasoning settings. Scores are relative to the current field, not a percentage of games solved.

View the full leaderboard

Programs scored by play

Models write player programs. DuelLab checks the code and runs matches between the working programs. The leaderboard reflects what those programs do in play, not how convincing the source code looks.

Public, evolving game set

Rankings cover named public games, with new games and evidence added over time.

Current, not frozen

GameBench 2 keeps changing, so use its latest standings as the current comparison.

A 30-second example

Prompt → code → match → score

  1. Prompt

    The model receives the game rules and instructions for sending moves.

  2. Check

    DuelLab checks whether the player program compiles, starts, and makes legal moves.

  3. Match

    Working player programs play recorded matches against each other.

  4. Score

    Wins, losses, draws, and the amount of evidence become a normalized 0–100 score.

Why this matters

Code that looks right can still fail.

Code can look correct and still fail to run, make legal moves, or play well. DuelLab builds each player program, checks it, and puts working programs into matches. The question is simple: does the program work, and how well does it play?

The result measures behavior in actual games rather than judging the generated code only by appearance or a small set of examples.

How it works

Behind the rankings

DuelLab uses the same basic process for every model. Here is how the public benchmark works.

  • Models write the players

    Each model receives the same written rules and move instructions for a game. It writes a player program, which DuelLab checks before match play begins.

  • Match results become scores

    Wins, draws, and losses produce a rating for each game. DuelLab then combines the game results and accounts for how much evidence supports them.

  • Public standings and games

    Rankings, game names, an updated date, and the methodology are public. GameBench 2 changes as new models, games, and match results are added or refreshed.

  • Models do not choose the games

    Games are selected and tested separately. Models may help test a game’s scope, but the benchmark task is to write players, not invent games.

Read the full methodology

Building next

From benchmark to lab.

Beyond the public rankings, DuelLab is building tools that describe games in reusable files instead of separate custom software. Those games can then be simulated many times while keeping their rules, replays, and experiments repeatable.

The longer-term goal is to build a public environment for inventing, testing, and analyzing games, with AI as an optional tool.