Four Gemini 4 Argon agents defeated Minecraft’s Ender Dragon in one reported survival run, according to the study’s authors. They played Minecraft Java Edition 1.20.4 on Easy, controlling the game through code and structured information about its world rather than by reading screen pixels. The result came from a single run, not a measure of repeatable success.

What the agents did

The four player agents entered a randomly generated world with cheat commands disabled and set out to defeat the Ender Dragon. From the first program to the reported win, the run took 27 hours and 14 minutes. The agents wrote 2,783 programs, exchanged 229 in-game chat messages and died 605 times. The dragon caused 69 of those deaths.

The authors also report that 22% of the programs ended in an error. Six server restarts paused the world for about 20 minutes in total. Across the full session, including activity after the win, 2,839 player programs ran.

How the agents controlled Minecraft

Each agent wrote JavaScript that ran through Mineflayer, an open-source bot library, and its pathfinding plugin. Instead of interpreting the game’s graphics and pressing keys, the agents read structured text about the game state and issued actions through their code. Mineflayer exposed information about the surroundings, including blocks behind solid rock, as well as details about the Ender Dragon’s internal state.

That setup matters to the result: it describes code-mediated play with access to structured game information, not gameplay through the same visual interface a person uses. The agents coordinated through in-game chat and shared persistent memory, with roles emerging during the run rather than being assigned beforehand.

A fifth agent, called the Observer, monitored the session but did not play. The researchers could watch a live dashboard and pause or stop the run. The authors say they did not intervene in gameplay. They recorded programs, outputs, chat and server logs; in their account, the logged “Free the End” advancement marked the win.

What the API estimates cover

The authors estimate that the four player agents’ API use through the win amounted to about $520 at introductory list prices or $1,040 at standard list prices. For the full session, including the Observer and activity after the win, their estimates rise to about $810 and $1,620, respectively.

Those figures are calculations based on token counters and list prices, not invoices or Google’s cost to serve the models. The players used about 1.35 billion input tokens through the win, with 91% served from context cache, and about 16 million output tokens, including 4 million thinking tokens. The authors also report roughly 24,000 model calls for the player agents through the win.

What one run establishes

The authors describe this as a single first-launch run on Easy difficulty. Its completion shows that this particular setup reached the game’s final objective; it does not establish a success rate across repeated attempts. The code-based interface, access to structured world information and the agents’ prior knowledge of Minecraft also define the conditions behind the result.