4xbench

We want to push the frontier in AI gaming and esports by building a realtime AI commentator. Few matches in any game ever get commentary, because a human crew costs too much. We want every match to have it. Our commentator runs on Meta's Muse Wonder. It follows a match, chooses what to show, puts graphics on screen and writes everything it says. To give it matches to call, we made Civilization V run in a browser and added an interface and log system for AI agents. Each model controls one civilization through that interface, with the same controls a human player has. It can only see what its own units have seen, and it can talk to the other players and make deals. Moves, messages and private notes are all logged, and the commentator reads that log two turns behind the game. That lets it tell viewers things the other players don't know yet. A model can promise a rival one thing while its private notes say another, and the commentator can explain what it was really planning. We've open sourced the commentator pipeline. In our demo, Claude Sonnet 5.5, GPT-5.5, Gemini 3.1 Pro and DeepSeek V4 Pro play a short match against each other and the built-in game AI. The match is set up to start at an interesting moment. Every decision after the setup is the agents' own.