Back to news
researchGoogle2026-11-11

Google and Artificial Analysis launch the Multimodal Knowledge Agent benchmark, with Gemini 4 Carbon ranking second

Google and Artificial Analysis launched the Multimodal Knowledge Agent benchmark with 200 tasks covering 10 industries and 83 occupations. Gemini 4 Carbon placed second, with average task cost roughly one sixth of the top-ranked model.

Google announced on November 10 that it had launched the Multimodal Knowledge Agent benchmark with the AI model evaluation organization Artificial Analysis, to assess how agents handle visually complex documents and complete real-world knowledge work tasks. Rohan Doshi, senior product manager at Google DeepMind, and chief engineer Ben Caine said many existing knowledge work benchmarks can be completed with text alone, while real jobs routinely involve substantial visual information, such as architects cross-checking building blueprints, engineers analyzing engineering diagrams, and professionals assembling information from different sources into presentations, reports and spreadsheets. The MKA evaluation set contains 200 tasks built with domain experts from 10 industries and 83 occupations, covering business, architecture, education, engineering, finance, healthcare, legal and government, logistics and mapping, and scientific research. Google said its newly released Gemini 4 Carbon took second place among participating models, completing tasks at an average cost of about one sixth of the first-ranked model. Carbon also leads several previously published benchmarks, including CursorBench 4.0, Perplexity WANDR and Zapier AutomationBench. The ARC Prize official leaderboard has also recorded Carbon validation results, listing an evaluation cost of 7200 dollars.

Artificial AnalysisGoogleGemini 4 CarbonMKA基准测试多模态