MKA (Artificial Analysis)

Google DeepMind's Multimodal Knowledge Agent (MKA) benchmark as run by Artificial Analysis, which holds its 200 private tasks: multi-step professional projects across 10 verticals, written with experts in 83 occupations, whose key information sits in visual files such as architectural drawings, engineering schematics, clinical scans, spreadsheets and slide decks. The agent gets a workspace of input files and must produce the deliverables a professional would (a report, spreadsheet, slide deck or marked-up drawing), working in Stirrup with a 250-turn limit, a Linux sandbox with code execution and web access, and an image-viewing tool for models with image input; each task runs once. A GPT-6 Luna (medium) agentic judge inspects the submitted files and final response and marks every expert-written rubric criterion, weighted 1 to 5, satisfied or not. Climb WSR, the headline, is the mean over tasks of the weighted share of the harder Climb criteria satisfied. Scores are for a model in Stirrup, not for the model alone.