Overview
- Google and evaluation partner Artificial Analysis made the Multimodal Knowledge Agent (MKA) benchmark public on Tuesday and published 10 sample tasks while keeping the benchmark’s 200 main tasks private to avoid data leakage.
- The MKA suite was built with experts across 10 industries and 83 professions to simulate multi‑step projects that require reasoning over blueprints, diagrams, slides and spreadsheets.
- Gemini 4 Carbon placed second on MKA’s quality ranking and completed tasks at about one‑sixth the average per‑task cost of the first‑place model, according to the initial results.
- Google has opened Gemini 4 Carbon to developers and enterprises via the Gemini API and Google AI Studio and is offering promotional pricing through November 30, 2026, with model specs that include support for text, images, video, audio and PDF and a unified 1,048,576‑token context window plus Long Decode output.
- To ensure fair comparison, all models ran in the same open‑source Stirrup agent framework with identical tools and sandboxing and were scored by a non‑Google automated reviewer whose outputs matched human expert scores about 96% of the time, and Google says Carbon shows strong security and research use cases while noting further safety work remains.