GPT-6.1 Sol beats Claude Opus 5.5 6-3, 6-1 in agent games

The experiments were run by Y Combinator Labs and Olam Labs, which built a custom open-source test environment letting agents observe the world, run Python scripts, and create files in a shared sandbox. In the Civ50 run, leadership emerged without a vote: 46 of the 49 other agents sent their first message to Ada, the first name in the directory, who then drafted settlement plans. After the first night's 22 deaths, Ada halted major optional construction, digging, and distant exploration, and average food levels settled at roughly half. A separate Minecraft run with the larger GPT-6 Astra showed a similar risk-averse potato-farming mode after losing key resources, which the researchers liken to coding agents becoming overly cautious around security or production updates.

The competitive rounds told a different story. In Bed Wars, GPT agents communicated 5 to 30 times more often than Claude opponents, often in compressed language, and in three games destroyed Claude's bed just before time expired. When an overthrown ender pearl teleported them roughly 160 blocks below the arena, they wrote Python scripts on the spot to build a tower up beneath the enemy bed, a maneuver observed twice. Across the Mindustry matches, Claude never swapped its leader and decided proposals through "+1" votes, while GPT split tasks into fine-grained pieces. Claude did show learning: after losing game six because 55 of its 74 anti-air turrets were loaded with copper instead of graphite, it built a ring of 207 graphite-loaded turrets in game seven and repelled the same air attack. The researchers argue that because the Mindustry setup avoided game-specific tuning and mirrors production coding environments, such simulations could surface dangerous behaviors before they appear in real systems.