Google and Artificial Analysis Launch MKA Benchmark
Google said it jointly launched the Multimodal Knowledge Agent benchmark, or MKA, with Artificial Analysis to assess AI agents on visually complex files and real-world knowledge-work tasks. [[fact:fact-fhf570h40]] The evaluation set contains 200 tasks and was developed with domain experts from 10 industries and 83 occupations. [[fact:fact-ks60j5k4z]] Google said Gemini 4 Carbon ranked second among participating models. [[fact:fact-740jqvb88]] Google said Gemini 4 Carbon’s average task-completion cost was about one-sixth that of the top-ranked model; this cost figure remains an unconfirmed estimate. [[fact:fact-73b1obzhr]] MKA is available through Artificial Analysis for developers to evaluate their own models. [[fact:fact-xqim33876]] The 200 formal evaluation tasks were kept private, while the project published 10 additional tasks on Hugging Face to show developers the test format and let them evaluate their own agents. [[fact:fact-z672nbs1b]][[fact:fact-o2sdqljs0]]