DuelLab · GameBench 2
We give AI models the rules of a game and ask them to write a program that plays it. We then run matches between the programs and rank the models by the results.
The model writes the code before the matches begin. During a game, the program makes every move without further help from the model.
GameBench 2
Results 2026-10-06
Showing the dated release snapshot.
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 leaderboardFigures from the 2026-10-06 public release.
One game, from rules to results
Chess is one of the public benchmark games. Every model's Chess program goes through the same three steps.
Explore the Chess resultsDoes the program compile, start, and make legal moves?
Working programs compete from recorded starting positions.
We estimate each player's strength from its match results, accounting for uncertainty. Overall scores also reflect how reliably each model produces working programs.
A selected recorded match · September 10, 2026
White: Muse Spark 1.2 · Medium
Black: OX Alpha · Max
Both sides are programs the models wrote before the match. Black wins by checkmate on move 41.
This selected example illustrates program play; it is not proof that the winning model is generally stronger.
September 10 source release. White program: E-2b12f5618349c056; Black program:
E-f1feaabe79bde601.
Download the recorded moves and board positions.
White's side is at the bottom. Coordinates run a1–h8.
Selected moment: 41…Rd4 delivers checkmate. Enable JavaScript to step through the recorded moves, or download the match record.
Why games?
Many benchmarks need a human or another model to grade the answers, so they can only measure skills the grader has. In a game, the result decides: a stronger program simply wins more often, however strong the models become. Read the argument →