Artificial Analysis and Google DeepMind Launch MKA Benchmark for Visual Agentic Work

Artificial Analysis partnered with Google DeepMind to launch the Multimodal Knowledge Agent (MKA) evaluation, a benchmark of 200 private real-world visual tasks across 10 industries requiring agents to produce deliverables such as slide decks, spreadsheets, and marked-up images. Claude Fable 5.2 leads the leaderboard with a Climb WSR of 57%, followed by Gemini 4 Carbon at 54% and Claude Opus 5.5 at 53%, with no single model leading across all 10 industry verticals.
Event version 1.
Key insights
Benchmark scope and task design
The MKA benchmark comprises 200 private real-world visual tasks across 10 industries, with agents working from dense visual project files including architectural drawings, engineering schematics, and clinical scans. Deliverables include slide decks, spreadsheets, PDFs, Word documents, and marked-up images.
Climb vs. Regression rubric scoring
Each task is graded against expert rubrics averaging about 70 weighted criteria, split into Regression baseline checks and harder Climb criteria. The headline score is Climb WSR, the weighted proportion of Climb rubric criteria successfully completed.
Evaluation harness and judge
Every task runs in Stirrup, an open source agent harness with code execution, web search, web fetch, and image-viewing tools up to 250 turns. Submitted deliverables are graded using GPT-6 Luna (medium) as an agentic judge.
Leaderboard results
Claude Fable 5.2 (max) leads with a Climb WSR of 57%, followed by Gemini 4 Carbon (high) at 54% and Claude Opus 5.5 (max) at 53%. No single model leads across all 10 industry verticals.
Cost-performance spread
Claude Fable 5.2 averages $31.97 per task, over 5× Gemini 4 Carbon at $6.29 for about 2 extra points, and about 29× GPT-6.1 Sol at $1.11. Below Carbon, all four OpenAI models tested sit on the cost-performance Pareto frontier, with Claude Haiku 5.5 and MiMo-V2.6-Pro making up the rest.