Four Gemini 4 Argon agents beat Minecraft autonomously in 27 hours, killing the Ender Dragon at $1,620 cost

Four AI agents running on Google's Gemini 4 Argon beat Minecraft, including the Ender Dragon, in a fully autonomous run with no human input. The Google DeepMind team says it is the first time this has been done without human help. The result comes from a paper by Faldor et al., reported by The Decoder on October 18, 2026.

The run in numbers

• Duration: 27 hours, 14 minutes
• Agents: 4, all running Gemini 4 Argon
• Programs written: 2,783 JavaScript programs
• Chat messages exchanged: 229
• Deaths: 605 across the four agents
• Cost: $1,620 at the standard Gemini 4 Argon price that takes effect in 2027

Setup

The DeepMind team placed the agents in a randomly generated Minecraft world. According to the report, they assigned no roles, no strategies and no game-specific tools.

The agents organized themselves. They split up tasks, warned each other about threats, recovered items from dead teammates, and logged mistakes in a shared memory. The agents acted by writing JavaScript programs, but the source does not describe the interface or scaffolding that executed them.

What is not disclosed

The source gives a total cost but no per-token pricing. Per-token pricing for Gemini 4 Argon is not yet disclosed here, and so are its context window and benchmark scores. The $1,620 figure is calculated at the price that takes effect in 2027, not at a rate the team necessarily paid. The "first without human input" framing is the team's claim as relayed by The Decoder, and we have not independently verified it.

Derived figures

These are our own calculations from the reported totals. They assume the 27h14m is wall-clock time:

• About $59.50 per hour for the four-agent team, or roughly $14.90 per agent-hour
• About 102 programs written per hour
• About 22 deaths per hour across the team

What this means

The 605 deaths show that the run was brute-force persistence, not clean play. The shared memory of logged mistakes is what let the team absorb that failure rate and still finish. For agent builders, memory and recovery design looks as important as raw model capability.

The cost sets a practical benchmark. At roughly $1,620 for one completion, long-horizon multi-agent runs are affordable for research but not for routine use. The figure also depends on a 2027 price that is not yet in effect.

The Decoder notes that games once requiring purpose-built systems trained from scratch are now being beaten by general-purpose agentic models. Whether the result generalizes beyond Minecraft depends on the paper's methodology, in particular what the JavaScript interface exposed to the agents. Until that is clear, treat this as a strong demonstration of multi-agent coordination, not proof of general autonomy.