Gemini Beat Minecraft in 27 Hours. The Setup and the Bill Matter.

Overhead view of the End island and obsidian pillars in the saved Minecraft research world.
The End arena from the researchers’ saved world. Image: Maxence Faldor, Satinder Singh and Roberta Raileanu, A team of Gemini 4 Argon agents beats Minecraft (CC BY 4.0). Pixel Perfection textures: XSSheep and Nova_Wostra (CC BY-SA 4.0). Image unchanged.


Defeating Minecraft’s Ender Dragon is an entertaining AI headline. What caught my attention was the less glamorous part: keeping a team working towards the same goal after hundreds of failures, and paying for all the attempts along the way.

For anyone interested in agents that can finish a long research or coding task, the useful question is how the result was achieved. A game victory is evidence of something, but we need to be clear about what.

The run, with its conditions attached

A Google DeepMind research paper reports four Gemini 4 Argon agents defeating the Ender Dragon in 27 hours and 14 minutes. They played Minecraft Java Edition 1.20.4 on Easy difficulty in a random world, with cheat commands disabled.

The team wrote 2,783 JavaScript programs, exchanged 229 messages and died 605 times before winning. A fifth agent observed. Researchers could stop the run but report no human intervention.

This was one run, not a reliability benchmark. The agents received structured game information and pathfinding support, rather than playing from screen pixels like a person. Prior Minecraft knowledge also mattered.

The cost of reaching the goal

The paper estimates about $520 for the four players through the win at introductory API list prices, or $1,040 at standard prices. The full session, including the observer and post-win activity, is estimated at $810 or $1,620 respectively.

Before the win, input usage was around 1.35 billion tokens, 91% cached. These are authors’ estimates from token counters and listed prices, not our measurements or provider production costs.

Why the interface changes the comparison

Mineflayer, the open-source JavaScript library used to control the players, provides a programmable way to interact with Minecraft. That is a different task from recognising the scene on a display and moving with a keyboard and mouse.

I do not see that as a reason to dismiss the result. Software agents often work through APIs and structured tools. For that kind of application, learning to coordinate those tools is precisely the interesting capability.

It does mean that comparing the headline time with a human speedrun would miss the point. Different interfaces, difficulty settings and team sizes make for different experiments. The stronger question is whether the system can preserve a useful plan, recover from errors and verify progress over a long sequence.

Persistence is valuable when failure is affordable

A game is a useful setting for exploring that behaviour because a failed attempt can be followed by another attempt. In an email account, an online shop or a production system, the consequences may be harder to undo.

That difference shapes what I want from a work agent. I want persistence when it is correcting a draft or debugging in an isolated environment. I want a pause when the next attempt requires new permissions, spending or an irreversible action.

The cost comparison is relevant too. A successful final result can hide a long sequence of retries. My post on completed-task costs makes the same point in a more everyday setting: the bill belongs to the whole process, not just its final answer.

What I want to see next

Repeated runs across different worlds would make this easier to assess. I would want to see the success rate, time and cost distribution, and the effect of removing particular tools. Those comparisons could show which parts of the system are doing the most useful work.

For now, I find this an interesting demonstration of coordination and persistence. It also pairs neatly with my earlier look at an AI taking a shortcut in a StarCraft task: a result becomes much more informative once we inspect the rules, the tools and the route to success.

Comments

Popular posts from this blog

Obsidian’s New Search: Less Friction for Notes in Two Languages

OPPO Find X9 Ultra vs Find X10 Pro Max: Is the Older Ultra Still Worth Buying?

OPPO Find X10 Technical Preview: Cameras, Design and Missing Specs