New MKA benchmark finds similar top AI scores at different task costs

Google DeepMind and Artificial Analysis launched MKA on November 10, a 200-task benchmark that tests AI agents on blueprints, schematics and scans, with most tasks requiring finished documents. Google's technical report found no statistically distinguishable headline-score difference among Claude Fable 5.2, Gemini 4 Carbon and Claude Opus 5.5. Google measured API costs of $35.28 per task for Fable 5.2 and $5.63 for Carbon. Removing Carbon's image-viewing tool cut its score by 5.94 percentage points while increasing token use. The headline score measures performance on harder grading criteria, rather than completed tasks; each model ran once per task, leaving run-to-run and grader variation outside the reported confidence intervals.