DeepMind's new AI test makes models read blueprints, not answer trivia

Google DeepMind and Artificial Analysis dropped a new benchmark on November 10, 2026, and it makes AI agents sweat over blueprints instead of trivia. The Multimodal Knowledge Agent benchmark, or MKA, arrived alongside a DeepMind blog post and a technical paper. It grades autonomous agents on dense, visually heavy professional work rather than puzzles.

MKA packs 200 private tasks spread across 10 industry verticals. Specialists from 83 different occupations helped shape them. Agents have to read architectural blueprints, engineering schematics and clinical scans, then hand back real deliverables: spreadsheets, slide decks, marked-up images. Each task mirrors multi-step projects that usually eat 3 to 8 hours of a human professional's day.

Scoring runs on strict expert rubrics that weight factual accuracy and visual design together. A right answer stuffed into a cluttered, unreadable deck still bleeds points. An agentic autorater does much of the judging, and per the research findings it agreed with human experts 96.2% of the time.

The first leaderboard carries a tidy result for the benchmark's co-author. Gemini 4 Carbon (high) landed second in quality — and finished tasks at roughly 6x lower cost than the model sitting in first place. DeepMind helped build the test and its own model scored well on cost-adjusted performance, a relationship worth keeping in mind when reading the numbers.

Keeping those 200 tasks private is deliberate: developers can't train their models on the answers beforehand. The benchmark targets a real gap, since most existing evaluations lean on text-only work while architects, engineers and clinicians all reason from drawings. With Artificial Analysis lending a third-party evaluation brand and DeepMind bringing research muscle and a competitive stake, whether other labs submit models and accept the methodology will decide if MKA becomes an industry standard.

The caveats are baked in. Two hundred tasks is hefty for a hand-curated benchmark, but it still only covers a slice of professional work, and strong results in one vertical may not carry over to another. The autorater's remaining disagreement with humans could also cluster in the hardest, most subjective tasks.