Minecraft · Minecraft

Can AI beat Minecraft? AI agents and the Ender Dragon

Minecraft has no fixed ending, so “beating Minecraft” usually means killing the Ender Dragon. Most AI agents play through code libraries such as Mineflayer, not with a screen and keyboard.

6 records and 1 tracked run or leaderboard for Minecraft only. Nothing on this page is compared with other games. How we verify.

Can AI beat Minecraft?

Reported, yes: Gemini 4 Argon: A team of 4 Gemini 4 Argon agents beat Minecraft: killed the Ender Dragon in 27h 14m, with 605 deaths and 2,783 programs written (creator-reported, Oct 15, 2026). Setup: harness: 4 general-purpose Gemini 4 Argon coding agents controlling players through Mineflayer (+ pathfinding plugin); team chat + shared memory; inputs: text API (position, health, inventory, nearby blocks and mobs), not pixels; human assistance: none during the run (researchers could pause or stop it but did not intervene); single run; Java Edition 1.20.4, survival, Easy difficulty, cheats off. The authors list the limits themselves: the agents read the world through an API rather than pixels, a plugin handled pathfinding, and it was a single run.

Records

Minecraft records
Model	Achievement	Run config	Status	Date
Claude Opus 4.6 Echo-Sinclair (OpenClaw)	#1 agent on KradleVerse: 25% win rate over the last 30 days Last verified: Oct 18, 2026 · Last checked: Oct 18, 2026 How was this run measured? Model Claude Opus 4.6 (source label: “claude-opus-4-6 (agent: Echo-Sinclair, OpenClaw)”) · agent: Echo-Sinclair (OpenClaw) Game Minecraft Project KradleVerse Harness OpenClaw agent framework Inputs not disclosed Tools not disclosed Assistance not disclosed Budget 30-day window, 10+ runs to qualify Verification Verified Read from the live operator leaderboard. Evidence KradleVerse leaderboard (official, live) (operator) Dates evidence dated Oct 18, 2026 Recorded published Oct 18, 2026 · discovered by AI vs Games Oct 17, 2026 Changes Last change 2026-10-18: 30-day win % 24% → 25% Permanent link · Submit a correction · Share card (PNG) · JSON	OpenClaw agent framework not disclosed	Verified source	Oct 18, 2026
Gemini 4 Argon Team of 4 Gemini 4 Argon agents	A team of 4 Gemini 4 Argon agents beat Minecraft: killed the Ender Dragon in 27h 14m, with 605 deaths and 2,783 programs written The authors call it the first end-to-end autonomous completion of Minecraft by AI agents. They also list its limits: the agents read the world through an API, a plugin handled pathfinding, and they played on Easy. Last verified: not independently verified · Last checked: Oct 18, 2026 How was this run measured? Model Gemini 4 Argon · agent: Team of 4 Gemini 4 Argon agents Game Minecraft Project Gemini 4 Argon agent team (research paper) Harness 4 general-purpose Gemini 4 Argon coding agents controlling players through Mineflayer (+ pathfinding plugin); team chat + shared memory Inputs text API (position, health, inventory, nearby blocks and mobs), not pixels Tools self-written JavaScript programs (2,783); no human-written game logic, no pre-scouted seed Assistance none during the run (researchers could pause or stop it but did not intervene) Budget single run; Java Edition 1.20.4, survival, Easy difficulty, cheats off Verification Creator-reported Reported by the paper's authors (submitted Oct 15, 2026). We have checked the paper; we have not seen independent replays or logs. Evidence Paper: 'A team of Gemini 4 Argon agents beats Minecraft' (arXiv 2610.22319) (creator) Full text (HTML) (creator) Unknown public replay or logs; compute cost Dates evidence dated Oct 15, 2026 Recorded published Oct 15, 2026 · discovered by AI vs Games Oct 18, 2026 Changes Historical record. Nothing to update unless the source changes. Permanent link · Submit a correction · Share card (PNG) · JSON	4 general-purpose Gemini 4 Argon coding agents controlling players through Mineflayer (+ pathfinding plugin); team chat + shared memory text API (position, health, inventory, nearby blocks and mobs), not pixels	Creator-reported source	Oct 15, 2026
GPT-6 Astra reasoning effort: medium	Top Mine Odyssey score: 85.6% of 180 navigation tasks in Minecraft rebuilds of real places Claude Opus 5.5 (max) was second at 73.9%. The best open-weight model managed 23.9%. Last verified: not independently verified · Last checked: Oct 18, 2026 How was this run measured? Model GPT-6 Astra (source label: “GPT-6 Astra (medium)”) Game Minecraft Project Mine Odyssey Harness Mine Odyssey navigation harness Inputs not disclosed Tools not disclosed Assistance none Budget reasoning effort: medium Verification Creator-reported Paper authors' results; no public leaderboard yet. Evidence Mine Odyssey paper (arXiv 2610.11328), Table 1 (creator) Dates achieved Oct 8, 2026 · evidence dated Oct 8, 2026 Recorded published Oct 8, 2026 · discovered by AI vs Games Oct 18, 2026 Changes Historical record. Nothing to update unless the source changes. Permanent link · Submit a correction · JSON	Mine Odyssey navigation harness not disclosed	Creator-reported source	Oct 8, 2026
GPT-6 Astra Astra + Jev Minecraft agent	GPT-6 Astra + Jev kill the Ender Dragon in 8m 43s on a fixed seed with a pre-surveyed route (Peaceful) Last verified: not independently verified · Last checked: Oct 18, 2026 How was this run measured? Model GPT-6 Astra (source label: “GPT-6 Astra (planner) + Jev (typesafe/jev-1.13)”) · agent: Astra + Jev Minecraft agent Game Minecraft Project Astra and JEV Minecraft agent Harness minecraft-agent (Mineflayer, fixed Nether route, read-only dragon sensor) Inputs structured game state; known seed coordinates Tools Mineflayer pathfinding, bounded scripted actions (travel, mine, craft, bed attack) Assistance no live operator help (creator-reported); pre-surveyed route on a fixed seed; Peaceful difficulty Budget $0.97 total ($0.01 Jev, $0.96 Astra) Verification Creator-reported Creator's post and README; the video's verification files are kept locally by the creator, not published. Evidence Creator's post: Jev + Astra beats the Ender Dragon in 8 minutes 43 seconds (Sep 20, 2026) (creator) Creator's code and README (run nether-final-08) (creator) Unknown number of attempts before nether-final-08; public full video link; Astra reasoning effort Dates achieved Sep 20, 2026 · evidence dated Sep 20, 2026 Recorded published Sep 20, 2026 · discovered by AI vs Games Oct 18, 2026 Changes Historical record. Nothing to update unless the source changes. Permanent link · Submit a correction · JSON	minecraft-agent (Mineflayer, fixed Nether route, read-only dragon sensor) structured game state; known seed coordinates	Creator-reported source	Sep 20, 2026
Claude Opus 5 Mine AI MCP agent (Claude Code)	Claude Opus 5 kills the Ender Dragon in a fresh Minecraft world in one unbroken 3h 56m session Last verified: not independently verified · Last checked: Oct 18, 2026 How was this run measured? Model Claude Opus 5 (source label: “claude-opus-5[1m] (effort high)”) · agent: Mine AI MCP agent (Claude Code) Game Minecraft Project Mine AI MCP Harness Claude Code + Mine AI MCP server (tag beat-the-game) Inputs MCP tool observations of the game world (no screenshots) Tools Mine AI MCP tools (navigation, mining, crafting, combat, exploration) Assistance none after the opening prompt (prompt includes strategy tips) Budget effort high, 1M context; $93.17 API-equivalent cost Verification Creator-reported Creator-published artifacts: we read the vanilla server log ('Free the End' at 04:39:50 UTC, no commands that affect the bot, one death) and the result file (Opus 5, effort high, 389 turns). Labelled creator-reported until we replay it ourselves; can be upgraded to verified after our own replay. Evidence Published run artifacts: transcript, tool-call database, vanilla server log, prompt (Hugging Face) (creator) Code: Mine AI MCP (tag beat-the-game) (creator) Full recording (YouTube) (creator) Unknown development sessions before this one; whether the prompt's strategy tips count as hints under other leaderboards' rules Dates started Sep 15, 2026 · achieved Sep 15, 2026 · evidence dated Sep 16, 2026 Recorded published Sep 16, 2026 · discovered by AI vs Games Oct 18, 2026 Changes Historical record. Nothing to update unless the source changes. Permanent link · Submit a correction · JSON	Claude Code + Mine AI MCP server (tag beat-the-game) MCP tool observations of the game world (no screenshots)	Creator-reported source	Sep 15, 2026
Claude Opus 4.6	Top MineExplorer score: 41.08 overall task success (open-world exploration) Last verified: not independently verified · Last checked: Oct 18, 2026 How was this run measured? Model Claude Opus 4.6 (source label: “Claude-Opus-4.6”) Game Minecraft Project MineExplorer Harness MineExplorer ReAct-style agent harness Inputs multimodal (screen + text) Tools not disclosed Assistance none Budget not disclosed Verification Creator-reported Paper authors' own results, listed as self-reported. Evidence BenchmarkList: MineExplorer (self-reported) (secondary) MineExplorer paper (creator) Dates evidence dated Jun 28, 2026 Recorded published Jun 28, 2026 · discovered by AI vs Games Oct 18, 2026 Changes Historical record. Nothing to update unless the source changes. Permanent link · Submit a correction · JSON	MineExplorer ReAct-style agent harness multimodal (screen + text)	Creator-reported source	Jun 28, 2026

Run history

Jun 28, 2026
Claude Opus 4.6: Top MineExplorer score: 41.08 overall task success (open-world exploration) Creator-reported
Sep 15, 2026
Claude Opus 5: Claude Opus 5 kills the Ender Dragon in a fresh Minecraft world in one unbroken 3h 56m session Creator-reported
Sep 20, 2026
GPT-6 Astra: GPT-6 Astra + Jev kill the Ender Dragon in 8m 43s on a fixed seed with a pre-surveyed route (Peaceful) Creator-reported
Oct 8, 2026
GPT-6 Astra: Top Mine Odyssey score: 85.6% of 180 navigation tasks in Minecraft rebuilds of real places Creator-reported
Oct 15, 2026
Gemini 4 Argon: A team of 4 Gemini 4 Argon agents beat Minecraft: killed the Ender Dragon in 27h 14m, with 605 deaths and 2,783 programs written Creator-reported
Oct 18, 2026
Claude Opus 4.6: #1 agent on KradleVerse: 25% win rate over the last 30 days Verified

Tracked runs & leaderboards

KradleVerse: AI agents play Minecraft LIVE

A live arena where community AI agents compete in Minecraft, ranked by 30-day win rate (minimum 10 runs).

How was this measured? (applies to every row)

Harness: community agents (framework shown per row)
Inputs: not disclosed
Tools: not disclosed
Assistance: not disclosed
Budget: ranked on last 30 days, 10+ runs to qualify
Sources: kradleverse.com

Submit a correction

#	Model	30-day win %	Status	As of
1	Claude Opus 4.6 agent: Echo-Sinclair (OpenClaw) 15 wins	25%	Verified	2026-10-18

Last checked: Oct 18, 2026 · auto-update: ok

This Week in AI vs Games

One short email every Sunday: new records, memorable failures and runs to watch. Free. Every claim is linked to its source.

Latest issue: Issue #1: The Ender Dragon is down