BenchLeader
Menu
Best LLMsChoose a modelCoding plansLeaderboardsModelsBenchmarksProvidersReleases7

BALROG leaderboard

As of 6 Nov 2026, the top score on BALROG is 68.3% by GPT-6 Astra (OpenAI), ahead of Claude Opus 5 at 63.4%, across 37 model configurations with a published result (BALROG, result dated 3 Sep 2026). Agents play games including BabyAI, Crafter, TextWorld and NetHack; the score is average progress toward each game's goal. BALROG is shown for reference only and does not count toward the BenchLeader Index.

Key facts (as of 6 Nov 2026)

  • GPT-6 Astra holds the top BALROG score at 68.3% (BALROG, 3 Sep 2026).
  • Claude Opus 5 is second on BALROG at 63.4% (BALROG, 24 Jul 2026).
  • BALROG: Progress through classic text and grid games such as NetHack (BALROG).
  • 37 model configurations have a published BALROG result on BenchLeader (BenchLeader, 6 Nov 2026).
  • BALROG does not feed the BenchLeader Index; it is shown for reference (BenchLeader methodology).

Progress through classic text and grid games such as NetHack.

Published by
BALROGdata via Epoch AI Benchmarking Hub
Category
Agents & tools
Index weight
Reference only
Models
37
Data as of
6 Nov 2026

CC BY 4.0 — Epoch AI, ‘AI Benchmarking Hub’, epoch.ai/benchmarks. Mirrored boards credit their original publishers.

What the test looks like

Agents play games including BabyAI, Crafter, TextWorld and NetHack; the score is average progress toward each game's goal.

How it is scored

Average progress across games, as published by BALROG.

What to keep in mind

Progress is normalised per game and games differ enormously in difficulty.

Top 15
  1. 1GPT-6 Astra (max)68.3%
  2. 2Claude Opus 5 (max)63.4%
  3. 3GPT-5.6 Sol (max)60.0%
  4. 4Gemini 3 Pro58.1%
  5. 5Gemini 3.1 Pro57.0%
  6. 6GPT-5.6 Terra (max)53.2%
  7. 7Gemini 3 Flash48.1%
  8. 8GPT-5.6 Luna (max)45.6%
  9. 9Grok 443.6%
  10. 10Gemini 2.5 Pro43.3%
  11. 11Claude Opus 4.543.0%
  12. 12DeepSeek R134.9%
  13. 13Gemini 2.5 Flash33.5%
  14. 14GPT-5 (minimal)32.8%
  15. 15Claude 3.5 Sonnet32.6%
37 of 37
#
1GPT-6 Astramax68.3%67.82026-09-03
2Claude Opus 5max63.4%65.42026-07-24
3GPT-5.6 Solmax60.0%65.42026-07-09
4Gemini 3 Pro58.1%60.22025-11-18
5Gemini 3.1 Pro57.0%60.52026-02-19
6GPT-5.6 Terramax53.2%60.72026-07-09
7Gemini 3 Flash48.1%56.92025-12-17
8GPT-5.6 Lunamax45.6%56.32026-07-09
9Grok 443.6%58.32025-07-09
10Gemini 2.5 Pro43.3%54.52025-03-25
11Claude Opus 4.543.0%57.92025-11-24
12DeepSeek R134.9%47.32025-01-20
13Gemini 2.5 Flash33.5%52.72025-06-17
14GPT-5minimal32.8%45.32025-08-07
15Claude 3.5 Sonnet32.6%47.12024-10-22
16GPT-4o32.3%43.32024-05-13
17Claude Haiku 4.531.2%47.62025-10-15
18Grok 329.5%52.52025-04-09
19Reka Flash 329.2%–2025-03-10
20Llama 3.1 70B27.9%40.82024-07-23
21Llama 3.2 90B Vision27.3%36.12024-09-24
22Llama 3.3 70B23.0%40.02024-12-06
23Gemini 1.5 Pro 00221.0%41.52024-09-24
24DeepSeek-R1-Distill-Qwen-32B19.5%–2025-01-20
25Claude 3.5 Haiku19.3%38.82024-10-22
26Mistral Nemo17.6%–2024-07-18
27GPT-4o mini17.4%35.42024-07-18
28Llama 3.2 Instruct 11B (Vision)16.8%35.72024-09-24
29Qwen2.5 72B16.2%43.22024-09-19
30Llama 3.1 8B15.1%33.72024-07-23
31Gemini 1.5 Flash 00214.6%38.22024-09-24
32Qwen2-VL-72B12.8%–2024-08-29
33Phi-411.6%39.42024-12-12
34Llama 3.2 3B Instruct10.1%35.82024-09-24
35Qwen2.5 7B Instruct7.8%39.82024-09-19
36Llama 3.2 1B Instruct6.6%30.72024-09-24
37Qwen2-VL-7B3.7%–2024-08-29

More agents & tools benchmarks · All LLM benchmarks

Related benchmarks

Top head-to-heads on BALROG

Cite as: BenchLeader, “BALROG leaderboard”, data as of 6 Nov 2026, https://www.benchleader.com/benchmarks/balrog.

BALROG: questions

What does BALROG measure?
Agents play games including BabyAI, Crafter, TextWorld and NetHack; the score is average progress toward each game's goal. Scores are reported in percent of tasks solved; higher is better.
What is the top score on BALROG?
As of 6 Nov 2026, the highest score on BALROG is 68.3% by GPT-6 Astra (result dated 3 Sep 2026), ahead of Claude Opus 5 at 63.4%. Scores come from BALROG and are re-read daily.
Which AI model leads BALROG?
GPT-6 Astra leads BALROG with 68.3% as of 6 Nov 2026 (result dated 3 Sep 2026), ahead of Claude Opus 5 at 63.4%.
How many models have BALROG results?
37 model configurations have a BALROG result on BenchLeader, all taken from BALROG via Epoch AI Benchmarking Hub.
Who runs BALROG and how often is it updated?
BALROG is published by BALROG. BenchLeader re-reads the published results once a day and records the date each result was published.
Does BALROG count toward the BenchLeader Index?
No. BALROG is shown for reference but left out of the composite index.