FrontierML ($50 wallet)
Agents · 2026-10-13
The same 21 FrontierML end-to-end machine learning tasks, harnesses, review and scoring as the $200 headline, with a $50 wallet per trial for GPUs and OpenRouter tokens: the mean over the 21 tasks of one trial's normalized score (0-1 between measured floor and ceiling anchors). Smaller wallets lower most scores, by about six points for Claude Opus 5.5, while a few agents score no better or worse with more money. It is a fallback component of the FrontierML factor group.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5.5 | 52.4 |
| Claude Sonnet 5.5 | 47.3 |
| GPT-6 Astra | 42.8 |
| Claude Fable 5.1 | 37.4 |
| GPT-6.1 Sol | 36.2 |
| Gemini 4 Argon | 30.4 |
| Claude Haiku 5.5 | 29.1 |
| Kimi K3.1 Pro | 26.7 |
| MiniMax M3.1 Flash | 24.0 |
| DeepSeek V4.1 Flash 1017 | 22.5 |
| Kimi K3 | 17.8 |
| GLM-5.3 | 17.4 |
| Muse Wonder | 17.2 |
| GPT-6 Luna | 13.3 |
| Muse Spark 1.3 | 12.6 |
| Grok 4.7 | 10.8 |
| Gemini 3.8 Flash | 9.5 |
| Beam 1.0 501B A23B | 9.3 |