Four Gemini 4 Argon agents defeated Minecraft’s Ender Dragon in one reported run, reaching the “Free the End” advancement 27 hours and 14 minutes after their first program, according to the paper’s authors. Their objective was to defeat the dragon, not to complete every activity in Minecraft’s open-ended world.
What the agents did in Minecraft
The run used Minecraft Java Edition 1.20.4 in survival mode, on Easy difficulty, in a randomly generated world with cheat commands disabled. The elapsed time runs from the agents’ first program to the game’s recorded advancement.
The four player agents exchanged 229 in-game messages and died 605 times before winning, including 69 deaths during the Ender Dragon fight. They wrote and ran 2,783 programs before the win; the paper separately counts 2,839 programs across the broader session.
How Mineflayer shaped the run
The agents controlled their players by writing JavaScript programs through Mineflayer, a library with a pathfinding plugin. They received text and structured game-state information, rather than controlling the game by interpreting screen pixels and pressing keys with a mouse and keyboard. The information included details such as blocks behind solid rock and the dragon’s internal state.
No one assigned the agents roles or gave them a route through the game. About ten minutes into the run, one agent proposed dividing the work. The group then shared discoveries and coordinates, specialized tasks and recovered teammates’ dropped items. At 13 hours and 17 minutes, they had activated the End portal; they later destroyed all ten End crystals before defeating the dragon.
The process involved substantial errors and interruptions: about 22% of the programs ended in an error, and the world server restarted six times, pausing the game for roughly 20 minutes in total.
What one run says about the result
The authors describe this as their first launch and report one successful run. That result does not provide a repeat-run success rate. It also describes code-mediated play through Mineflayer under the stated Easy-mode conditions, not visual gameplay using pixels, a keyboard and a mouse.
The paper’s novelty claim is narrower than “the first AI to beat the Ender Dragon.” The authors qualify it as, to their knowledge, the first end-to-end autonomous Minecraft completion by AI agents using standard Mineflayer, without pre-scouted seeds or game-specific tools written by humans. Their related-work discussion acknowledges earlier AI-related Ender Dragon kills under different conditions.
What the API cost estimates cover
The authors estimate API usage at public list prices using token counters. They put the four player agents’ usage through the win at about $520 at introductory prices or $1,040 at standard prices. For the broader session, including the non-playing Observer agent and a day of activity after the win, their estimates are about $810 or $1,620, respectively.
| Scope | Introductory list-price estimate | Standard list-price estimate |
| Four player agents through the win | About $520 | About $1,040 |
| Full session, including the Observer and a day after the win | About $810 | About $1,620 |
Through the win, the authors count about 1.35 billion input tokens, with 91% served from context cache, roughly 16 million output tokens, including 4 million thinking tokens, and about 24,000 model calls. They describe the dollar figures as list-price estimates based on those counters, not model-serving costs.
