Google says Gemini 4 Argon agents beat Minecraft without human help

Google has announced that its latest AI model, Gemini 4 Argon, has successfully completed the game of Minecraft without any human intervention. In a research paper released on October 15, DeepMind researchers detailed how a group of four Gemini 4 Argon agents managed to defeat the Ender Dragon, the game's final boss, in a freshly generated survival world on Easy difficulty.

The AI players completed the task in 27 hours and 14 minutes, despite dying 605 times in the process, primarily due to monster attacks. This marks the first time AI agents have completed Minecraft from start to finish using only standard tools, without any human assistance or game-specific modifications. The researchers observed the run through a live dashboard and could have intervened at any point, but decided not to.

However, Google is clear about the limitations of this accomplishment. The game was played on Easy difficulty, which reduced the damage dealt by monsters, and the AI agents had access to certain information that human players wouldn't, such as knowing the layout of blocks. The AI agents also had a training base that informed them of optimal strategies for playing the game.

While this is an impressive feat, it's still far from the capabilities of human players. The researchers compare this to the record for beating Minecraft in a random world, which is under seven minutes. They question the relevance of this benchmark, as Minecraft is often used to test AI agents' abilities to work together, recover from mistakes, and persist in a task for extended periods.

The researchers estimate that the cost of running this experiment was about $520, using Google's introductory API price for the Argon model. This type of benchmark is common among tech companies when testing their frontier AI models. While Pokémon games have been a popular choice for such tests in the past, Minecraft is gaining traction as well.

The research highlights two community projects that have recently reported achieving a similar feat with other AI models, including one using Anthropic’s Claude Opus 5.