Introducing the Multimodal Knowledge Agent benchmark

Google and Artificial Analysis have launched the Multimodal Knowledge Agent benchmark, a new evaluation set designed to measure how AI agents reason across visually complex files and real-world workflows. The benchmark consists of 200 multi-step tasks across 10 verticals created with domain experts, and is graded using an agentic autorater. Gemini 4 Carbon achieved the second-highest quality score in testing while completing tasks at a significantly lower cost. Open article →