RTSGameBench - Duel 1v1: leaderboard
RTSGameBench (Kim et al., 2026) full-game duel in the real-time strategy game Beyond All Reason: the vision-language model plays one side through the RTSGameAgent harness (FSM group management plus agentic memory, the same VLM for memory and decisions) against the built-in AI at Easy difficulty on a fixed map with fog of war. Win rate over 30 games, in percent; higher is better.
Metric: Win Rate (%; 30 runs vs Easy built-in AI). Source: arxiv.org. Saturation forecast: Around February 2027. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash | 87 | #159 |
| 2 | GPT-5.2 | 53 | #172 |
| 3 | Grok 4.1 Fast | 33 | #426 |
| 4 | Kimi K2.5 | 30 | #252 |
| 5 | Claude Sonnet 4.5 | 27 | #244 |
| 6 | Qwen 3 VL 235B A22B (Thinking) | 20 | #451 |
| 7 | Qwen 3.5 397B A17B | 13 | #242 |
| 8 | GPT-5 Mini | 7 | #349 |
| 9 | Qwen 3 VL 235B A22B Instruct | 0 | #521 |
| 10 | Mistral Large 3 | 0 | #819 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=rtsgamebench-duel-1v1 · How It Works · Data refreshed daily, snapshot 2026-11-12.