More benchmarks
Beyond the main event
Benchmarks we still track, but which don't (yet) have live, dependable sources or a big fan following. NetHack moves to the main scoreboard once a dependable live source exists. Know one?
Records
| Model | Game | Achievement | Status | Date |
|---|---|---|---|---|
| Gemini 2.5 Pro (preview 03-25) | Kirby's Dream Land | Memorable fail 4.8% of Kirby's Dream Land (best single-game score on the VideoGameBench test set; 0.48% overall) Last verified: Oct 18, 2026 · Last checked: Oct 18, 2026 How was this run measured?
|
Verified | May 23, 2025 |
| GPT-6 Astra reasoning effort: max |
NetHack | Memorable fail 13.2% progress on NetHack for the top overall BALROG entry. Nobody has beaten the game. Last verified: Oct 18, 2026 · Last checked: Oct 18, 2026 How was this run measured?
|
Verified | Sep 18, 2026 |
Leaderboards
BALROG: BabyAI, Crafter, TextWorld, BabaIsAI, MiniHack, NetHack RECENT
Average progress across 6 game environments. NetHack is still the hardest. The leaderboard updates weekly on Mondays.
How was this measured? (applies to every row)
- Harness
- BALROG agentic harness
- Inputs
- text / vision per track
- Tools
- none
- Assistance
- none
- Budget
- not disclosed
- Sources
- balrogai.com
| # | Model | Avg progress % | Status |
|---|---|---|---|
| 1 | GPT-6 Astra (max) NetHack 13.2% |
68.3 | Verified |
| 2 | Claude Opus 5 (max) NetHack 7.3% |
63.4 | Verified |
| 3 | GPT-5.6 Sol (max) NetHack 3.2% |
60 | Verified |
| 4 | Gemini 3 Pro NetHack 6.8% |
58.1 | Verified |
| 5 | Gemini 3.1 Pro (thinking) NetHack 2.6% |
57 | Verified |
| 6 | Gemini 3.1 Pro NetHack 3.0% |
56.9 | Verified |
| 7 | GPT-5.6 Terra (max) NetHack 3.3% |
53.2 | Verified |
| 8 | Gemini 3 Flash NetHack 4.0% |
48.1 | Verified |
| 9 | GPT-5.6 Luna (max) NetHack 1.7% |
45.6 | Verified |
| 10 | Grok 4 NetHack 1.8% |
43.6 | Verified |
Last checked: Oct 18, 2026 · auto-update: ok
BalatroBench ARCHIVE
Ranks LLMs by the average round they reach in Balatro when playing through tool calls. The live table loads client-side, so the current standings are not verified.
⚠ Leaderboard data generated 2026-02-03 (benchmark version v1.0.8).
How was this measured? (applies to every row)
- Harness
- BalatroBench default strategy
- Inputs
- game state (text)
- Tools
- not disclosed
- Assistance
- none
- Budget
- not disclosed
- Sources
- balatrobench.com · github.com
| # | Model | Avg final round | Status |
|---|---|---|---|
| 1 | Gemini 3 Pro Preview 9/15 runs won |
19.3 ± 6.8 | Verified |
| 2 | GPT-5.2 3/16 runs won |
18.4 ± 5.1 | Verified |
| 3 | Gemini 3 Flash Preview 5/15 runs won |
16.7 ± 7.4 | Verified |
| 4 | Claude Opus 4.5 2/15 runs won |
13.9 ± 6.5 | Verified |
| 5 | Claude Sonnet 4.5 2/15 runs won |
11.9 ± 6.7 | Verified |
| 6 | Grok 4 0/15 runs won |
9.5 ± 4.9 | Verified |
| 7 | DeepSeek-V3.2 0/15 runs won |
8.7 ± 6.8 | Verified |
| 8 | Grok 4.1 Fast 0/15 runs won |
7.7 ± 5.5 | Verified |
| 9 | Qwen3-Max 0/15 runs won |
7.3 ± 4.3 | Verified |
| 10 | gpt-oss-120b 0/15 runs won |
6.3 ± 2.9 | Verified |
Last checked: Oct 18, 2026 · auto-update: ok
Factorio Learning Environment ARCHIVE
Agents write programs to build factories and are scored by production. The data is stale: only older models are listed, and the official page currently shows zeros.
⚠ Stale: no current-generation models listed.
How was this measured? (applies to every row)
- Harness
- Factorio Learning Environment (code API)
- Inputs
- text / code
- Tools
- FLE Python API
- Assistance
- none
- Budget
- not disclosed
- Sources
- jackhopkins.github.io · benchmarklist.com
| # | Model | Production score | Status |
|---|---|---|---|
| 1 | Claude 3.5 Sonnet | 293,206 | Verified |
| 2 | Gemini 2.0 Flash | 115,782 | Verified |
| 3 | GPT-4o | 87,599 | Verified |
| 4 | Llama 3.3 70B | 54,998 | Verified |
| 5 | DeepSeek-V3 | 48,585 | Verified |
| 6 | GPT-4o mini | 26,756 | Verified |
Last checked: Oct 18, 2026 · auto-update: failed, showing last good data
Orak (KRAFTON AI): 12 video games ARCHIVE
Covers Street Fighter III, Super Mario, Ace Attorney, Pokémon, Minecraft, Stardew Valley, StarCraft II, Slay the Spire and more. Last updated Jun 9, 2025.
How was this measured? (applies to every row)
- Harness
- Orak benchmark harness
- Inputs
- per game
- Tools
- not disclosed
- Assistance
- not disclosed
- Budget
- not disclosed
- Sources
- krafton-ai.github.io
| # | Model | Avg rank (lower is better) | Status |
|---|---|---|---|
| 1 | Gemini 2.5 Pro (preview 03-25) | 3.9 | Verified |
| 2 | GPT-4o (2024-08-06) | 4.3 | Verified |
| 3 | o3-mini | 4.6 | Verified |
Last checked: Oct 17, 2026 · auto-update: manual source
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