Google DeepMind introduced the Multimodal Knowledge Agent benchmark on November 10 with Artificial Analysis. It evaluates deliverables such as reports, spreadsheets and presentations across 200 private tasks. Ten public examples do not make the full evaluation set public.

What the result measures

Artificial Analysis runs agents in a Linux sandbox with code, search and image tools and a 250-turn limit. Its headline Climb Weighted Success Rate averages the weighted share of harder rubric criteria satisfied per task. Baseline Regression criteria are separate.

The November 12 page shows Claude Fable 5.2 at Max with Default Fallback scoring 57%, Gemini 4 Carbon at High 54%, and Claude Opus 5.5 at Max with Default Fallback 53%. Effort settings are part of each result.

Attribution matters

Google helped develop the benchmark; Artificial Analysis describes its own implementation and independent runs. This is not a wholly vendor-independent design or a percentage of jobs agents can complete.

These scores are neither Intelligence Index nor Terminal-Bench. They do not establish that evaluated Carbon High matches the separately released public API model.