Arena 用 4700 场 agent 会话评测 Jev Router

David Hendrickson @TeksEdge原推发布于 2026.10.08

在 X 上打开原推

Curious Image. Does your local AI need a router or just a good default model? Arena evaluated Jev Router across 4,700+ real-world agentic sessions. It reports performance similar to DeepSeek V4.1 Flash (Max), but ... 💰 38% higher cost ⏱️ 1.7× higher median model-request latency Its strength was responding to user corrections, where its steerability score nearly matched Opus 5.5. Jev selects other LLMs to do the work. To clarify, this isn’t a small model matching Opus by itself. 🎯The question is: Does building routing into your Agents actually improve the completed tasks enough to justify the extra overhead?