DeepMind Unveils MKA Benchmark to Evaluate AI Agents on Professional Visual Tasks

Google DeepMind, in collaboration with Artificial Analysis, has launched the Multimodal Knowledge Agent (MKA) benchmark to evaluate AI agents on real-world professional visual tasks. The benchmark comprises 200 private, real-world tasks across 10 industries, requiring agents to interpret complex visual inputs such as architectural blueprints, engineering diagrams, and clinical scans, and to produce deliverables like slide decks, tables, and reports. It fills a gap in existing evaluations for professional scenarios like blueprint cross-referencing and scientific chart reading.

MKA employs a dual-track expert scoring system (Climb and Regression) with an average of about 70 weighted criteria per task, emphasizing authentic visual complexity and domain specificity over simplified synthetic data. Results show Claude Sonnet 5.5 leading in areas such as healthcare, though its advantage in legal and government domains is marginal. Gemini 4 Carbon ranks second, with an average cost per task six times lower than the top model, demonstrating significant cost-effectiveness.

The launch marks a shift in multimodal agent evaluation from general capabilities to vertical professional scenarios, providing developers with a tool that closely mirrors real workflows. As AI agents penetrate knowledge work, MKA is poised to become an industry standard, driving improvements in visual understanding, cross-domain reasoning, and deliverable generation, while influencing enterprise adoption and cost-benefit trade-offs.