Google Multimodal Knowledge Agent (MKA) Benchmark
In brief
Google and Artificial Analysis released MKA, a benchmark evaluating how AI agents perform multi-step, visually complex tasks across 10 professional verticals.

Benchmarks, how it works and examples
3 from 1 source
BenchmarkChart
How it worksDiagram
ExampleIllustrationHow it works
Each task pairs a prompt and multimodal input files with an expert rubric of around 70 checks. Models run inside the open-source Stirrup agent harness with identical tools and turn budgets. An automated agent evaluates deliverables on factual accuracy and visual design.
AI researchers and developers evaluating multimodal agent performance on real-world visual tasks.
Availability
Sources
2 merged into this cardMore from Google
Gemini CLI /stats heatmap subcommand
Google added a /stats heatmap subcommand to Gemini CLI v0.69.0 to track daily token usage. It lets developers monitor their token consumption directly from the command line.
Gemini CLI adds standalone theme.json support
Google's gemini-cli v0.69.0 adds support for a standalone theme.json configuration file. Users can now define custom themes in a dedicated file.
