Frontier-Bench v0.1
Agents · 2026-07-22
Frontier-Bench v0.1 evaluates general-purpose agents on 74 difficult tasks across seven domains using rich terminal environments, GPU and multi-container workloads, open-internet access, and programmatic verifiers. Scores report mean reward for a named model, agent harness, and reasoning setting.
Top models (higher is better)
| Model | Score |
|---|---|
| Claude Opus 5 | 44.4 |
| GPT-5.6 Sol | 37.5 |
| Claude Fable 5 | 33.8 |
| Opus 4.8 | 21.1 |
| GPT-5.6 Terra | 20.8 |
| Grok 4.5 | 17.8 |
| Sonnet 5 | 17.0 |
| GPT-5.6 Luna | 14.3 |
| GLM-5.2 | 5.1 |