Orak - StarCraft II: leaderboard
Orak StarCraft II: an LLM agent plays a full match against the built-in AI through 72 discrete macro actions (training, building, research); score is the win rate over trials, 0–100, with the game's default agentic strategy. Higher is better; few trials make the score coarse. From KRAFTON AI's Orak benchmark (NeurIPS 2025).
Metric: Game Score (0-100). Source: krafton-ai.github.io. Saturation forecast: Estimated already saturated. 12 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro (Preview 03-25) | 100 | #126 |
| 2 | GPT-4o (2024-08-06) | 100 | #343 |
| 3 | GPT-4o Mini | 75 | #598 |
| 4 | DeepSeek R1 | 50 | #262 |
| 5 | Claude 3.7 Sonnet | 50 | #259 |
| 6 | O3 Mini (2025-01-31) | 25 | #249 |
| 7 | Qwen 2.5 7B Instruct | 0 | #863 |
| 8 | Llama 3.2 3B Instruct | 0 | #1327 |
| 9 | Llama 3.2 1B Instruct | 0 | #1518 |
| 10 | Qwen 2.5 3B Instruct | 0 | #1159 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=orak-starcraft-ii · How It Works · Data refreshed daily, snapshot 2026-11-05.