Google DeepMind and Artificial Analysis introduced the Multimodal Knowledge Agent (MKA) benchmark on November 10, 2026. The project evaluates autonomous AI agents on complex visual tasks, such as interpreting architectural blueprints, engineering schematics, and clinical scans.
The benchmark consists of 200 private tasks across 10 industry sectors, developed with input from specialists in 83 occupations. These tasks simulate professional projects that typically require 3 to 8 hours for a human to complete.
MKA utilizes an agentic autorater to score outputs based on factual accuracy and visual design quality. Research indicates this system achieves a 96.2% agreement rate with human experts.
In initial results, Google's Gemini 4 Carbon (high) model secured the second position in quality. The model demonstrated efficiency by completing tasks at a cost approximately 6 times lower than the top-ranked model.
