Google's Multimodal Knowledge Agent benchmark: Claude Fable 5.2 leads at 57%

The benchmark was built to fill a gap: existing agent tests are too text-heavy. As the Google blog authors Rohan Doshi and Ben Caine put it, most knowledge work benchmarks can be solved with text alone, without ever needing to look at an image or page layout, yet everyday professional work looks very different. MKA packages 200 tasks across 10 verticals and 83 occupations, with each task reflecting a multi-step project that takes a seasoned professional three to eight hours to complete.

All models run inside the same open-source agent harness, Stirrup, with a shared toolset and turn budget. Even so, the tested models show extreme differences in efficiency, using between 30 and 130 turns on average to complete each task. The autorater's 96% agreement with domain experts offers some confidence in the scoring, though the private task set and Google's co-creation role leave open questions about independence that Crypto Briefing has flagged.

The result places Google in an unusual spot: its Gemini 4 Carbon ranks second on a benchmark it helped design, behind Claude Fable 5.2. The post does not name the model that beat Gemini 4 Carbon on quality, and this is the third Google alert on the tracker today, following the wider Gemini 4 Carbon API rollout.