Google launches MKA benchmark for AI agents

Google has teamed up with AI testing firm Artificial Analysis to launch MKA, a new benchmark that measures how well AI agents handle visually complex files and real-world knowledge work.

The test was built with experts from 10 industries and 83 jobs. It includes 200 tasks covering business, architecture, healthcare, law and more. Each task is a multi-step project that usually takes a professional 3 to 8 hours, with about 70 check items per task.

Results show Google's new Gemini 4 Carbon ranked second in quality, while its average cost per task was only about one-sixth of the top model. Human expert scores matched the automated grading system 96% of the time.

MKA is now open through Artificial Analysis. The 200 official tasks are kept private to prevent leaks, while 10 tasks are public on Hugging Face for developers to test their own agents.