Google DeepMind launches MKA, a benchmark for agents working from visually complex files

Artificial Analysis says it partnered with Google DeepMind on the Multimodal Knowledge Agent (MKA) evaluation, a benchmark for agentic knowledge work over visually complex project files: engineering schematics, architectural drawings and clinical scans. Agents must use them to produce slide decks, spreadsheets, reports and marked-up images. The tasks span 10 verticals, and a 10-task public set is available.

Artificial Analysis reports that no single model dominates across verticals. Claude Fable 5.2 leads five of them (Software & Data, Engineering, Logistics & Mapping, Business, and Finance, the last alongside Claude Sonnet), and leads four of the six output file types (PDF, Word, narrowly Spreadsheet, and Presentation alongside Claude Opus 5.5 (max)).

On one example task, it says Gemini 4 Carbon (high) finished in 76 turns and about an hour for roughly $8, against 224 turns, about three hours and roughly $68 for Claude Fable 5.2 (max). Fable 5.2's deck scored higher (87% vs 72%). Rubric criteria are split into "Climb" (harder, where frontier models still often fail) and "Regression" (baseline competence); on Regression the leading seven models all score 95% to 97%.

These are Artificial Analysis's own figures; Google DeepMind published its own launch blog for the benchmark.