Agents working from schematics and scans are scored on expert rubrics, with cost per task reported. Scores come within about 10 points at roughly 30x cost spread, so model choice for visual document work is a cost decision too.
Key takeaways · AI-distilled
Google DeepMind's MKA benchmarkA standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.Full definition → has 200 private tasks across 10 industries in which agents must read dense visual files such as architectural drawings, schematics and clinical scans, then produce slide decks, spreadsheets, documents or marked-up images.
Each MKA task is graded against expert rubrics averaging about 70 weighted criteria, split into baseline 'Regression' checks and harder 'Climb' checks. The headline score, Climb WSR, is the weighted share of Climb criteria met.
Artificial Analysis runs every task in its open-source Stirrup agent harnessThe scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.Full definition → with code execution, web search, web fetch and image viewing for up to 250 turns, and uses GPT-6 Luna (medium) as an agentic judge of the deliverables.
Claude Fable 5.2 (max) leads at 57% Climb WSR, but at $31.97 per task it costs over 5x Gemini 4 Carbon (54%, $6.29) and about 29x GPT-6.1 Sol ($1.11), according to Artificial Analysis.
No model leads every vertical: Fable 5.2 leads five, while Opus 5.5, Gemini 4 Argon, Gemini 4 Carbon and Sonnet 5.5 lead the others. Word document and text file outputs are the hardest, with no model above 52% on either.
Terms in this piece · Glossary
multimodal — A model that works with more than text — reading images, audio, or video, and sometimes generating them too.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
