Orak Battle Arena - StarCraft II: leaderboard
Orak Battle Arena for StarCraft II: seven zero-shot LLM agents play pairwise Protoss-versus-Protoss matches, one round per pair, and the outcomes are summarized as Elo ratings (higher is better). With a single game per pairing the ratings are coarse. From KRAFTON AI's Orak benchmark (NeurIPS 2025).
Metric: Elo. Source: krafton-ai.github.io. Saturation forecast: Estimated already saturated. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.7 Sonnet | 1550.1 | #259 |
| 2 | GPT-4o (2024-08-06) | 1525 | #343 |
| 3 | GPT-4o Mini | 1500 | #598 |
| 4 | O3 Mini (2025-01-31) | 1500 | #249 |
| 5 | Qwen 2.5 7B Instruct | 1445.2 | #863 |
| 6 | Llama 3.2 3B Instruct | 1424.9 | #1327 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=orak-battle-arena-starcraft-ii · How It Works · Data refreshed daily, snapshot 2026-11-05.