Gemini 4 Argon Beats Minecraft After 27 Hours of Agent Chaos

Gemini 4 Argon beat Minecraft without human intervention. Four agents defeated the game’s final Ender Dragon foe after working individually, sharing tips, threats, and mistakes along the way.

The experiment took 27.23 hours, turning a familiar game into a prolonged test of planning, coordination, and recovery. The agents wrote and ran 2,783 programs, exchanged 229 messages, and died 605 times before reaching the finish.

They also used 1.35 billion input tokens. At Gemini’s list price, the final cost was $810; at the standard price, it was $1,620. Nothing says progress like making four agents spend a day failing at the same dragon until the accounting department notices.

A game victory built from thousands of decisions

Google researchers said the agents needed to work together without a single agent controlling the entire process. Their shared messages carried practical knowledge, warnings, and failures between separate attempts, allowing the group to maintain progress across the experiment.

“The four agents wrote and ran 2,783 programs, exchanged 229 messages and died 605 times along the way,” the team said. Those numbers show the shape of the achievement: less a clean march to victory than a long chain of experiments, setbacks, and adjustments.

“Winning required the agents to keep a coherent plan across thousands of dependent steps for more than a day,” the team said. Beating the Ender Dragon therefore tested more than isolated commands; it required the agents to preserve a plan while handling the consequences of earlier decisions.

Google DeepMind developed Gemini 4 Argon for tasks involving finance, software engineering, coding, creative writing, and cybersecurity defense. The company said the model matches OpenAI GPT-6 Astra’s score on the Intelligence Index while costing 60 percent as much per task at its introductory prices.

Argon moves into Google’s consumer products

Google first rolled out Gemini 4 Argon to members of its Fairwind Program earlier in October 2026. It is now available to Google AI Ultra and Google AI Pro subscribers through the Gemini app and AI Mode in Search.

Google AI Pro costs $20 a month, while Google AI Ultra costs at least $100. Access to Argon costs $2 per million input tokens and $10 per million output tokens during the introductory period, but those prices will double after November 30, 2026.

Google had been expected to launch a 3.5 Pro model but released Argon instead. The model’s broader role now covers demanding work across finance, software engineering, coding, creative writing, and cybersecurity defense—an ambitious list, even by the standards of companies that describe every new model as ready for everything.

Inside the Gemini app, Argon answers more quickly and uses less hedging, fewer repetitions, and fewer clichés. Josh Woodward, VP of Product for Gemini and Google Labs, said, “In the Gemini app, Argon answers more quickly and with less hedging, repetitions and clichés.”

Google is also making it easier to toggle between chat and tasks within the Gemini app. That change puts conversation and agent work closer together, reducing the separation between asking for an answer and asking the system to carry out a task.

The company’s macOS app will add Antigravity’s agentic coding interface inside Gemini. Alongside the Minecraft demonstration, the addition points toward a product strategy built around agents that write programs, run them, exchange information, and continue working through failure.

Argon’s Minecraft result is striking because the agents did not win through one perfect move. They won after 2,783 programs, 229 messages, 605 deaths, 1.35 billion input tokens, and 27.23 hours of persistence. The dragon lost eventually. The invoice did not.

Stay connected via Google News

Follow us for the latest AI News