Gemini 4 AI Agents Beat Minecraft in 27 Hours, Costing $810

Four autonomous AI agents running on Google’s Gemini 4 model defeated Minecraft’s final boss without human help, racking up 605 deaths and an $810 compute bill over 27 hours. The run marks the first time general-purpose artificial intelligence has completed the survival game end-to-end using only standard open-source tools, according to a new paper published by Google researchers.

Previous attempts to build AI that plays video games relied on custom software trained specifically for a single title. DeepMind’s AlphaGo, which beat world champions at the board game Go, required specialized architecture. This experiment used off-the-shelf models operating through Mineflayer, an open-source library for interacting with Minecraft servers. No pre-scouted seeds. No game-specific code written by humans.

The setup dropped four agents — identified in the logs as “janesick,” “gyldenia,” “auratian” and “bulicame” — into a randomized sandbox survival world. Their objective: kill the Ender Dragon. They received no assigned roles or strategies. Instead, they coordinated through in-game chat and shared memory.

They figured out the division of labor themselves. “janesick” focused on crafting tools. “gyldenia” gathered food and supplies. “auratian” located and lit the portal required to reach the final dimension. “bulicame” built shelters to manage hostile Enderman mobs and delivered the killing blow to the dragon.

The coordination went beyond simple task delegation. The agents warned each other about incoming threats, exchanged coordinates, offered gameplay tips and retrieved inventory from teammates who died. A fifth agent, designated “the Observer,” watched the entire session independently. Its job was to verify the others weren’t cheating and to generate live reports, the researchers said.

The numbers behind the victory are stark. The four agents wrote and executed 2,783 programs. They exchanged 229 messages. The process consumed 1.35 billion input tokens. At Gemini’s list price, the total cost came to $810; at the standard rate, it hit $1,620.

Most of the 27.23 hours were spent fighting the Ender Dragon itself. “Winning required the agents to keep a coherent plan across thousands of dependent steps for more than a day,” the team wrote in the paper.

Minecraft has long served as a testing ground for AI research because its open-ended mechanics force systems to handle planning, resource management and unpredictable environments simultaneously. Microsoft, which owns the game, released Project Malmo in 2016 specifically to let researchers train AI inside the platform. But those earlier efforts generally produced agents capable of isolated tasks — chopping wood, navigating mazes — rather than completing the full game loop autonomously.