Google DeepMind's MKA Benchmark Forces AI Agents to Read Blueprints Like Professionals

Google DeepMind and Artificial Analysis launched MKA , a 200-task benchmark for visually complex agent work Tasks span 10 verticals and 83 occupations, each needing 3-8 hours of expert time Claude Fable 5.2 leads at 57% Climb WSR, Gemini 4 Carbon second at 54%, Claude Opus 5.5 third at 53% Cost per

Source

Artificial Analysis

Published

Nov 10, 2026, 6:22 PM

Author

Artificial Analysis

Read

1 min read

Takeaways

Google DeepMind and Artificial Analysis launched MKA , a 200-task benchmark for visually complex agent work
Tasks span 10 verticals and 83 occupations, each needing 3-8 hours of expert time
Claude Fable 5.2 leads at 57% Climb WSR, Gemini 4 Carbon second at 54%, Claude Opus 5.5 third at 53%
Cost per task spans nearly 30x: Fable at $32, Carbon at $6, GPT-6.1 Sol at $1.11
No model wins all verticals; Gemini leads Science and Healthcare, Claude dominates most others
Public 10-task set on Hugging Face , runs in the open-source Stirrup harness

Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.

Read original report