Gemini learns to inspect video selectively

Google DeepMind announced agentic video understanding for several Gemini Flash models. Instead of treating every part of a clip identically, the feature lets the model inspect segments dynamically. Google reports reduced token usage and cost alongside quality improvements in its evaluations.

Context

Selective inspection can be useful when a question concerns a brief event inside a long recording. It also makes coverage an evaluation issue: the system must not skip the moment that changes the answer. A good result should identify the relevant segment so a person can verify it directly.