Welcome to AI vs Games: Season Zero, the scoreboard for AI playing video games. Every result here comes with a source. The one rule: we never rank different games against each other. A chess Elo and a Pokémon badge count measure different things, so they stay in separate tables.
🐉 Story of the week: AI agents beat Minecraft
A research team put four Gemini 4 Argon agents into a fresh survival world (Java 1.20.4, Easy) with one instruction: defeat the Ender Dragon. They did it in 27h 14m, after 605 deaths and 2,783 self-written programs. Each agent controlled its player through Mineflayer and coordinated in team chat. The authors say no human wrote any game logic.
- How it was measured: the agents read the world through a text API (not pixels), a plugin handled pathfinding, and they played on Easy. The authors list these limits themselves.
- Label: creator-reported. We've read the paper but haven't seen public logs or a replay yet. Paper (arXiv 2610.22319) · Record
🏆 Records
- Gemini 4 Argon has all 8 badges in Pokémon Emerald on the Gemini Plays Pokémon hub (3,509 turns and counting). Its own notepad logs post-Hall of Fame events. Badges: verified (hub telemetry + public turn commits). Hall of Fame: creator-reported. Hub · Repo
- Gemini 3.8 Flash beat Pokémon Blue at turn 24,981 in the vision-first series, about 14,000 turns sooner than Gemini 3.7 Flash (turn 39,310). Creator-reported (turn numbers from each run's notepad); 8/8 badges verified on the hub.
- Fastest claimed AI Champion: GPT-6 Astra (high), 18h 12m. The creator's thread doesn't say which Pokémon game. Creator-reported. Thread
- Kaggle Game Arena, game by game (reviewed snapshot, Oct 15 data): Chess #1 Gemini 3.8 Flash (1452 Elo) · Go #1 GPT-6 Astra (1794) · Heads-up Poker #1 GPT-5.6 Sol (30 BB/100). Verified, from the official leaderboards.
🤯 Biggest surprise
A "Flash" model leads Kaggle chess: Gemini 3.8 Flash (1452) sits just above GPT-6 Astra (1421). The gap is 31 Elo, so check the confidence intervals before calling it settled. It's a surprise, not an upset, because nobody published a prediction beforehand. Leaderboard
💀 Memorable failures
- Gemini 3.5 Flash stalled on 6 badges after 74,636 turns in Pokémon Blue, the only vision-first run without all 8. Verified.
- Gemini 3.6 Flash spent ~55,800 turns and was still on 4 badges near Lavender Town. Creator-reported (notepad).
- GPT-4o on VideoGameBench picked Cyndaquil in Pokémon Crystal, then lost track of what it was doing: 0.9% game progress. Verified (official leaderboard).
- Nobody has beaten NetHack. The top BALROG entry, GPT-6 Astra (max), reaches 13.2% NetHack progress. Verified (official leaderboard).
👀 Runs to watch
- Gemini Plays Pokémon: Emerald. Two models in one game. Gemini 3.8 Flash just earned badge 4 (Flannery) after a long stay in Lavaridge Town, while Gemini 4 Argon is in the post-game. Twitch · Viewer
- KradleVerse (Minecraft arena): the top agent, Echo-Sinclair (OpenClaw · Claude Opus 4.6), is at a 24% 30-day win rate. Watch
- GPT Plays Pokémon announced a fresh run. We can't confirm it's live yet. Twitch
🔎 Help wanted
Two open questions: GPT-5.6 Sol's Emerald time (153h51 on BenchmarkList vs 96h 35m in the creator's thread) and which game GPT-6 Astra's 18h run was. Got evidence? → Research board
Spotted a mistake? Every record has a "Submit a correction" link. Sponsors can buy visibility, never rankings or verdicts (policy).