RSI Arena Stage 1 human rating
Agents · 2026-10-24
Whether an AI agent can train a model people prefer. Ten agents each train the same base model, Nemotron 3.5 Lightning, for up to 144 hours with $300 of API credit and 1,000 GPU-hours, and people then compare the resulting models in blind side-by-side battles. The score is the trained model's style-controlled human rating, credited to the agent that trained it. Unrelated to Vals' RSI Index.
Benchmark source: Rsiarena
Top models (higher is better)
| Model | Score | Source |
|---|---|---|
| GPT-6 Astra | 1095 | RSI Arena paper Table 3 (GPT-6 Astra, rating) · Bake AI |
| Grok 4.7 | 1038 | RSI Arena paper Table 3 (Grok 4.7, rating) · Bake AI |
| MiMo-V2.6-Pro | 1025 | RSI Arena paper Table 3 (MiMo-V2.6-Pro, rating) · Bake AI |
| GLM-5.3 | 1024 | RSI Arena paper Table 3 (GLM-5.3, rating) · Bake AI |
| MiniMax M3.1 Flash | 1002 | RSI Arena paper Table 3 (MiniMax M3.1 Flash, rating) · Bake AI |
| Kimi K3 | 982 | RSI Arena paper Table 3 (Kimi K3, rating) · Bake AI |
| Gemini 3.8 Flash | 981 | RSI Arena paper Table 3 (Gemini 3.8 Flash, rating) · Bake AI |
| Muse Spark 1.3 | 938 | RSI Arena paper Table 3 (Muse Spark 1.3, rating) · Bake AI |
| Step 5 Preview | 934 | RSI Arena paper Table 3 (Step 5 Preview, rating) · Bake AI |
| DeepSeek V4.1 Flash | 887 | RSI Arena paper Table 3 (DeepSeek V4.1 Flash, rating) · Bake AI |