Gemini 3.5 Flash
6
6 of 8 badges in Pokémon Blue (74,636 turns, idle)
Verified recordLATEST
- Oct 18, 2026 1:49 PM ET · Gemini Plays Pokémon: Emerald (Gemini 4 Argon & Gemini 3.8 Flash): Gemini 3.8 Flash Badges 4 → 5
- Oct 18, 2026 12:49 PM ET · KradleVerse: AI agents play Minecraft: Claude Opus 4.6 30-day win % 24% → 25%
- Oct 18, 2026 6:51 AM ET · KradleVerse: AI agents play Minecraft: Claude Opus 4.6 30-day win % 23% → 24%
- Oct 18, 2026 6:51 AM ET · Gemini Plays Pokémon: Emerald (Gemini 4 Argon & Gemini 3.8 Flash): Gemini 3.8 Flash Badges 3 → 4
- Oct 17, 2026 11:48 PM ET · ClaudePlaysPokemon: Red (Claude Opus 4.7): Claude Opus 4.7 Live now yes → no
- Oct 17, 2026 11:45 PM ET · ClaudePlaysPokemon: Red (Claude Opus 4.7): Claude Opus 4.7 Live now no → yes
★ SEASON ZERO
Where AI meets the final boss
The scoreboard for AI playing video games
Can AI beat Pokémon, Minecraft or chess? We track every run and leaderboard one game at a time, and every result links to its source.
Updated Oct 18, 2026 2:05 PM ET
Season Zero totals
40sourced records
19verified from primary evidence
13games on the board
18creator-reported
Biggest moment
AI agents beat Minecraft
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.
Why it's here: A claimed end-to-end Ender Dragon kill by AI agents that wrote all their own game code: no human-written game logic, no pre-scouted seed, no human help during the run. Other Minecraft dragon kills on our board use human-built tools or pre-planned routes. Reported in a research paper and labelled creator-reported until independent logs or replays are public.
Creator-reported Evidence: Paper: 'A team of Gemini 4 Argon agents beats Minecraft' (arXiv 2610.22319) · Full text (HTML)
Full record & how it was measured →
Last verified: not independently verified · Last checked: Oct 18, 2026
Scoreboard
Agent Adventures
- Creator-reported Gemini 4 Argon · Minecraft: 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
- Verified Gemini 4 Argon · Pokémon Emerald: All 8 badges in Pokémon Emerald (now 3,722 turns in, still playing)
- Verified Gemini 3.8 Flash · Pokémon Emerald: 5 of 8 badges in Pokémon Emerald after 8,825 turns (live)
- Creator-reported GPT-6 Astra · Pokémon (game not stated): Became Champion in 18h 12m, the fastest claimed AI Champion time we know of
Official Benchmarks
- Verified Gemini 3.8 Flash: #1 on Kaggle Game Arena Chess: 1452 Elo
- Verified GPT-6 Astra: #1 on Kaggle Game Arena Go: 1794 Elo
- Verified Heatz agent (model not disclosed): 1st: reached the first gym (Roxanne) in 40:13 (1,608 actions)
- Verified Claude Opus 4.6: #1 agent on KradleVerse: 25% win rate over the last 30 days
Live & Notable Runs
LIVE Gemini Plays Pokémon: Emerald (Gemini 4 Argon & Gemini 3.8 Flash)
Two Gemini models are playing Pokémon Emerald on the Gemini Plays Pokémon hub. Gemini 4 Argon has all 8 badges after 3,722 turns, and Gemini 3.8 Flash has 5 badges after 8,825 turns. Gemini 3.8 Flash was last seen at Route 119.
LIVE KradleVerse: AI agents play Minecraft
A live arena where community AI agents compete in Minecraft, ranked by 30-day win rate (minimum 10 runs).
RECENT Gemini Flash 'vision-first' series: Pokémon Blue
A series of 'vision-first' Pokémon Blue runs with successive Gemini Flash models on the same harness. Gemini 3.8 Flash: 8 badges; Gemini 3.7 Flash: 8 badges; Gemini 3.5 Flash: 6 badges. Each run's notepad on the public repo names the game as Pokémon Blue.
Biggest surprise
A "Flash" model leads the Kaggle chess ladder
Gemini 3.8 Flash (1452 Elo) sits above GPT-6 Astra (1421 Elo) on Kaggle Game Arena text chess, where models play without engine help. The gap is only 31 Elo, so check the confidence intervals on Kaggle before calling it settled.
Why it's here: Flash is Google's name for its faster, lighter tier, so seeing it at #1 is surprising. Nobody published a pre-match prediction, so we call this a surprise, not an upset.
Verified Evidence: official Kaggle leaderboard (as of Oct 15, 2026) · Chess page →
Then vs Now
Gemini Flash generations in Pokémon Blue
Metric: Badges (out of 8), from hub game-state telemetry · same game, same 'vision-first' run series on the same harness
⚠ Harness version per run is not documented. We treat these runs as comparable because the hub labels them as one series.
Gemini 3.7 Flash
8
8 of 8 badges in Pokémon Blue (55,601 turns, idle)
Verified recordGemini 3.8 Flash
8
8 of 8 badges in Pokémon Blue (47,889 turns, idle)
Verified recordTurns to Champion: 3.7 Flash vs 3.8 Flash
Metric: Turns to Champion (harness turns) · same game, same series; turn numbers taken from each run's notepad (creator-reported)
⚠ Turn length may differ between runs if the harness changed. Not documented.
Latest moves full log →
- Gemini Plays Pokémon: Emerald (Gemini 4 Argon & Gemini 3.8 Flash): Gemini 3.8 Flash Badges 4 → 5
- KradleVerse: AI agents play Minecraft: Claude Opus 4.6 30-day win % 24% → 25%
- KradleVerse: AI agents play Minecraft: Claude Opus 4.6 30-day win % 23% → 24%
- Gemini Plays Pokémon: Emerald (Gemini 4 Argon & Gemini 3.8 Flash): Gemini 3.8 Flash Badges 3 → 4
- ClaudePlaysPokemon: Red (Claude Opus 4.7): Claude Opus 4.7 Live now yes → no
- ClaudePlaysPokemon: Red (Claude Opus 4.7): Claude Opus 4.7 Live now no → yes
How we score: one game at a time. We never merge different games into a single “smartest AI” ranking. Every result shows the exact model version, the game, the run setup (harness, inputs, tools, assistance, budget, or “not disclosed”), the achievement, a source, a verification label and when we last checked. Methodology →
Can AI beat video games? Questions people ask
Can AI beat video games?
Some, depending on the setup. Finished games on our board: 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); Gemini 3.8 Flash: became Champion of Pokémon Blue at turn 24,981 (Hall of Fame at turn 24,994) (creator-reported, Sep 27, 2026); GPT-5: became Champion of Pokémon Red in 161h57m (6,470 steps, 8 badges) (creator-reported, Aug 14, 2025). Others are nowhere near done: GPT-6 Astra: 13.2% progress on NetHack for the top overall BALROG entry. Nobody has beaten the game (verified, Sep 18, 2026); Gemini 2.5 Pro (preview 03-25): 4.8% of Kirby's Dream Land (best single-game score on the VideoGameBench test set; 0.48% overall) (verified, May 23, 2025). How far an AI gets depends a lot on its harness, tools and any help, so we score every game separately.
Which AI is best at video games?
No honest single answer exists, because each game measures something different. We rank one game at a time. On Kaggle Game Arena right now: chess: Gemini 3.8 Flash (1,452 Elo, as of Oct 15, 2026); Go (board game): GPT-6 Astra (1,794 Elo, as of Oct 15, 2026); poker: GPT-5.6 Sol (30 Mean BB/100, as of Oct 8, 2026). Pick a game below to see its own leaderboard.
How do AI models play video games?
Through a harness: a program that shows the model the game (screenshots, game-memory data or a text description) and turns its replies into button presses or code. Some agents never look at a screen. The Minecraft team behind the Ender Dragon kill acted through code (4 general-purpose Gemini 4 Argon coding agents controlling players through Mineflayer (+ pathfinding plugin); team chat + shared memory). Every record lists its harness, inputs, tools and human assistance, or says “not disclosed”.
Can AI beat Pokémon?
Yes, in several games and setups. See every AI Pokémon run, or go straight to Red, Blue, Emerald or Crystal.
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