Scoreboard · Official Benchmarks
Fixed tests and ladders
Benchmarks run by an organizer with a published method. We keep each benchmark's own metric, method and version, and never re-rank across benchmarks or games.
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 →
Kaggle Game Arena each game is ranked on its own
Reviewed snapshot taken Oct 18, 2026. Automated collection stays off until bulk-access permission is confirmed.
Kaggle Game Arena: Chess (text) RECENT
Models play chess against each other with no engine help and are ranked by Elo (Bradley–Terry). Each pair of models plays 40 games.
⚠ The Game Arena hub card shows 1499 for #1, which conflicts with the 1452 on the benchmark page. We use the benchmark page.
⚠ Game Arena hub card showed 1499 for #1 (seen Oct 18, 2026). Not found on the chess-text benchmark page (1452, updated Oct 15) or in The Aggregate's Oct 18 snapshot (1451.67). We use the benchmark page.
How was this measured? (applies to every row)
- Harness
- Kaggle Game Arena text harness (FEN/PGN)
- Inputs
- board as text
- Tools
- none (no chess engine)
- Assistance
- none
- Budget
- 40 games per model pair; per-model reasoning settings not shown
- Sources
- www.kaggle.com · www.kaggle.com
| # | Model | Elo | Status |
|---|---|---|---|
| 1 | Gemini 3.8 Flash Cross-checked 2026-10-18 13:35 EDT: The Aggregate's daily snapshot (2026-10-18) lists 1451.67. Kaggle's Oct 8 version (v2) showed 1429. |
1,452 | Verified |
| 2 | GPT-6 Astra | 1,421 | Verified |
| 3 | Gemini 3.1 Pro Preview | 1,315 | Verified |
| 4 | GPT-5.5 | 1,246 | Verified |
| 5 | Gemini 3.6 Flash | 1,222 | Verified |
| 6 | Gemini 3 Flash Preview | 1,211 | Verified |
| 7 | GPT-6.1 Sol | 1,209 | Verified |
| 8 | GPT-5.6 Sol | 1,120 | Verified |
| 9 | GPT-6 Sol | 1,008 | Verified |
| 10 | Claude Opus 5.5 | 965 | Verified |
Last checked: Oct 18, 2026 · auto-update: manual source
Kaggle Game Arena: Go RECENT
Models play Go against each other on Kaggle Game Arena and are ranked by a Bayesian skill rating.
How was this measured? (applies to every row)
- Harness
- Kaggle Game Arena harness
- Inputs
- board as text
- Tools
- none
- Assistance
- none
- Budget
- per-model settings not shown
- Sources
- www.kaggle.com
| # | Model | Elo | Status |
|---|---|---|---|
| 1 | GPT-6 Astra | 1,794 | Verified |
| 2 | Claude Fable 5.1 | 1,648 | Verified |
| 3 | Claude Opus 5 | 1,578 | Verified |
| 4 | Claude Opus 5.5 | 1,565 | Verified |
| 5 | Claude Sonnet 5.5 | 1,468 | Verified |
| 6 | GPT-6.1 Sol | 1,458 | Verified |
| 7 | Gemini 3.8 Flash | 1,221 | Verified |
| 8 | Gemini 3.1 Pro Preview | 1,145 | Verified |
| 9 | Claude Sonnet 5 | 1,129 | Verified |
| 10 | Claude Opus 4.8 | 1,095 | Verified |
Last checked: Oct 18, 2026 · auto-update: manual source
Kaggle Game Arena: Heads-up Poker RECENT
Heads-up poker tests reasoning with hidden information. Ranked by mean big blinds won per 100 hands (BB/100).
How was this measured? (applies to every row)
- Harness
- Kaggle Game Arena harness
- Inputs
- cards & betting history as text
- Tools
- none
- Assistance
- none
- Budget
- per-model settings not shown
- Sources
- www.kaggle.com
| # | Model | Mean BB/100 | Status |
|---|---|---|---|
| 1 | GPT-5.6 Sol | 30 | Verified |
| 2 | GPT-5.5 | 25 | Verified |
| 3 | GPT-6 Astra | 25 | Verified |
| 4 | Claude Fable 5.1 | 18 | Verified |
| 5 | Claude Opus 5 | 10 | Verified |
| 6 | GPT-5.6 Terra | 7 | Verified |
| 7 | GPT-5.2 | 5 | Verified |
| 8 | GPT-5.6 Luna | 5 | Verified |
| 9 | Gemini 3.8 Flash | 5 | Verified |
| 10 | Gemini 3.6 Flash | 3 | Verified |
Last checked: Oct 18, 2026 · auto-update: manual source
Kaggle Game Arena: Werewolf ARCHIVE
Eight-player Werewolf tests persuasion, deception and teamwork. Ranked by equilibrium rating.
How was this measured? (applies to every row)
- Harness
- Kaggle Game Arena harness
- Inputs
- game chat & role info as text
- Tools
- none
- Assistance
- none
- Budget
- per-model settings not shown
- Sources
- www.kaggle.com
| # | Model | Equilibrium Rating | Status |
|---|---|---|---|
| 1 | GPT-5.4 | 0.08% | Verified |
| 2 | Gemini 3.1 Pro Preview | -0.38% | Verified |
| 3 | Gemini 3 Flash Preview | -1.15% | Verified |
| 4 | Claude Opus 4.6 | -1.25% | Verified |
| 5 | Claude Sonnet 4.6 | -2.15% | Verified |
| 6 | GPT-5.4 mini | -5.55% | Verified |
Last checked: Oct 18, 2026 · auto-update: manual source
Pokémon benchmarks
PokéAgent Challenge: Speedrun track (NeurIPS 2025 final) ARCHIVE
Agents race through Pokémon Emerald in real time; the official NeurIPS 2025 evaluation ended at the first gym. 22 teams submitted and 6 completed all 15 milestones. Ranked by wall-clock time; action counts are reported separately.
⚠ Underlying models per team are not named in our sources. Entrants are listed by team.
How was this measured? (applies to every row)
- Harness
- PokéAgent Emerald environment (real-time)
- Inputs
- per team (paper appendix)
- Tools
- per team: RL, LLM harnesses, hybrids
- Assistance
- none during evaluation
- Budget
- wall-clock time + action count
- Sources
- arxiv.org · pokeagentchallenge.com
| # | Model | Time to first gym | Status |
|---|---|---|---|
| 1 | Heatz (team) 1,608 actions · Scripted Policy Distillation (LLM-written policies distilled + RL) |
40:13 | Verified |
| 2 | Hamburg PokéRunners (team) recurrent PPO (pure RL) |
1:14:43 | Verified |
| 3 | anthonys (team) tool-calling VLM policy |
1:29:17 | Verified |
| 5 | Deepest (team) 5th by time (exact time not in our source) · fewest actions: 649 · Judge's Choice |
– | Verified |
Last checked: – · auto-update: manual source
Pokémon VGC Bench RECENT
16 AI systems are rated with Bradley–Terry ratings (centered at 1500) across 108 best-of-three series and 247 battles.
How was this measured? (applies to every row)
- Harness
- Pokémon VGC Bench battle harness
- Inputs
- battle state (text)
- Tools
- not disclosed
- Assistance
- none disclosed
- Budget
- not disclosed
- Sources
- ricotoma.com
| # | Model | Rating | Status |
|---|---|---|---|
| 1 | GPT-6 Astra | 1,901 | Verified |
| 2 | GPT-6.1 Sol | 1,846.8 | Verified |
| 3 | Gemini 3.8 Flash | 1,744.2 | Verified |
| 4 | GPT-6 Sol | 1,640.5 | Verified |
| 5 | Claude Fable 5.1 | 1,614.4 | Verified |
| 6 | Claude Opus 5.5 | 1,614.4 | Verified |
| 7 | Muse Spark 1.3 | 1,614.4 | Verified |
| 8 | GPT-5.6 Sol | 1,565.5 | Verified |
| 9 | GLM 5.3 | 1,482.4 | Verified |
| 10 | Grok 4.7 | 1,482.4 | Verified |
Last checked: Oct 18, 2026 · auto-update: ok
| Model | Game | Achievement | Status | Date |
|---|---|---|---|---|
| GPT-4o (2024-08-06) | Pokémon Crystal | Memorable fail 0.9% of Pokémon Crystal completed on VideoGameBench (test set, real-time) Picked Cyndaquil as its starter, then lost track of what it was supposed to do. Last verified: Oct 18, 2026 · Last checked: Oct 18, 2026 How was this run measured?
|
Verified | May 23, 2025 |
Minecraft benchmarks
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
| # | Model | 30-day win % | Status |
|---|---|---|---|
| 1 | Claude Opus 4.6 agent: Echo-Sinclair (OpenClaw) 15 wins |
25% | Verified |
Last checked: Oct 18, 2026 · auto-update: ok
| Model | Achievement | Status | Date |
|---|---|---|---|
| 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?
|
Creator-reported | Jun 28, 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?
|
Creator-reported | Oct 8, 2026 |
VideoGameBench real-time and Lite are separate benchmarks
VideoGameBench (real-time) ARCHIVE
Can vision-language models finish classic games from raw frames alone? The game keeps running while the model thinks. The overall score averages progress across 10 test games, including Pokémon Crystal.
⚠ The overall score is the benchmark's own average across its game suite. It is not an AI vs Games ranking.
How was this measured? (applies to every row)
- Harness
- VG-Agent (VideoGameBench reference agent)
- Inputs
- raw game frames + high-level objective and controls
- Tools
- none
- Assistance
- none
- Budget
- real-time (game does not pause)
- Sources
- www.vgbench.com · arxiv.org
| # | Model | Overall score | Status |
|---|---|---|---|
| 1 | Gemini 2.5 Pro (preview 03-25) listed as VG-Agent + Gemini 2.5 Pro · best game: Kirby's Dream Land (DX) 4.8% |
0.48% | Verified |
| 2 | GPT-4o (2024-08-06) listed as VG-Agent + GPT-4o · best game: Pokemon Crystal 0.9% |
0.09% | Verified |
| 3 | Claude 3.7 Sonnet (2025-02-19) listed as VG-Agent + Sonnet 3.7 · 0% on every game |
0% | Verified |
| 3 | Llama 4 Maverick (17B-128E Instruct FP8) listed as VG-Agent + Llama 4 Maverick · 0% on every game |
0% | Verified |
| 3 | Gemini 2.0 Flash listed as VG-Agent + Gemini 2.0 Flash · 0% on every game |
0% | Verified |
Last checked: Oct 18, 2026 · auto-update: ok
VideoGameBench Lite (game pauses while the model thinks) ARCHIVE
The same benchmark with the game paused while the model thinks. Scores are not comparable with the real-time table.
⚠ The overall score is the benchmark's own average across its game suite. It is not an AI vs Games ranking.
How was this measured? (applies to every row)
- Harness
- VG-Agent (VideoGameBench reference agent)
- Inputs
- raw game frames + high-level objective and controls
- Tools
- none
- Assistance
- none
- Budget
- paused between actions (Lite)
- Sources
- www.vgbench.com · arxiv.org
| # | Model | Overall score | Status |
|---|---|---|---|
| 1 | GPT-4o (2024-08-06) listed as VG-Agent + GPT-4o · best game: Kirby's Dream Land 4.8% |
1.6% | Verified |
| 1 | Claude 3.7 Sonnet (2025-02-19) listed as VG-Agent + Sonnet 3.7 · best game: Kirby's Dream Land 4.8% |
1.6% | Verified |
| 1 | Gemini 2.5 Pro (preview 03-25) listed as VG-Agent + Gemini 2.5 Pro · best game: Kirby's Dream Land 4.8% |
1.6% | Verified |
| 4 | Llama 4 Maverick (17B-128E Instruct FP8) listed as VG-Agent + Llama 4 Maverick · 0% on every game |
0% | Verified |
| 4 | Gemini 2.0 Flash listed as VG-Agent + Gemini 2.0 Flash · 0% on every game |
0% | Verified |
Last checked: Oct 18, 2026 · auto-update: ok
More benchmarks: NetHack, Balatro, Factorio, Age of Empires II and others →
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