Last week we launched agentic video understanding in Gemini! It navigates timelines dynamically instead of ingesting every frame, cutting tokens by up to 88% and costs by up to 66%! it also supports YouTube links. you can test it in Google AI Studio (toggle it on in settings):

note that it is not enabled by default in the API. You can enable it by setting processing to "agentic". we also wrote a developer guide:

@GoogleAIStudio gemini found the skip intro button

@GoogleAIStudio Works like magic! Tested it in AI Studio and it’s incredible. Will we be able to use this in AntiGravity as well?

@GoogleAIStudio @grok does it still look at all the frames?

@GoogleAIStudio Does Antigravity support video understanding (both video attachments and YouTube links)? Would be really cool

@GoogleAIStudio I am testing using my youtube channel video link

@GoogleAIStudio I cant see where it is. Google always does a gazillion products no one knows off🤦‍♂️ no wonder most of them get shutdown

@GoogleAIStudio This available via the CLI as well?

@GoogleAIStudio Add it in antigravity as well as aistudio pleeeeeeease

@GoogleAIStudio That is a useful cost lever, but the benchmark should include retrieval misses and temporal jumps, not just token savings. If the model skips the decisive frame, an 88% reduction is a cheaper wrong answer.

@GoogleAIStudio The agy harness and rules around using the subscription with other harness are doing a disservice to your model

@GoogleAIStudio 88% fewer tokens but only 66% cheaper is the interesting part. the seeking itself costs something, and whatever tokens survive are the expensive ones

The 66% cut is the headline. The open question is coverage. If the model skips a stretch, it can answer fluently without opening the decisive seconds. Fine for a creative prompt. Not fine for evidence or compliance. Wrote this up with the investing read (Alphabet vs Axon, Samsara, Motorola):

@GoogleAIStudio Cutting 88% of tokens is the real flex. Frame by frame video analysis gets expensive fast.

@GoogleAIStudio I love it! I'm building an app that creates linear issues from videos of usability tests for apps.

@GoogleAIStudio add this to antigravity

@GoogleAIStudio Thank you Gemini this is the one area of AI you are the undisputed king!

@GoogleAIStudio I’m just waiting for Google to dominate AI in 2027, I have a feeling

@GoogleAIStudio wow such a dumb approach

@GoogleAIStudio Does the Google cloud agent platform support agentic video understanding?

@GoogleAIStudio Can it also give the transceitp as it is?

@GoogleAIStudio Can you ask gemini for speech to text for video and then post it 😭

@GoogleAIStudio @patloeber @GoogleAIStudio Any issue with gemini-3.1-pro-preview? In AI Studio web (incl. Build) & Gemini API Tier 3, thinking and generation are faster, but quality is lower and words are fewer. Suspect output is routed to Lite. If it is a bug, please check and fix!

@GoogleAIStudio 88% is significant success. And I think you can go beyond that.

@GoogleAIStudio The token cut is the useful part. Skipping irrelevant frames makes long video analysis feel like retrieval instead of brute force.

@GoogleAIStudio This is huge for robotics training

@GoogleAIStudio is this active in gemini/antigravity ?

@GoogleAIStudio Is the feature available in Antigravity IDE?

@GoogleAIStudio how about in Agy :D

@GoogleAIStudio Is there a length cap? I have a 1 and 2 hour videos I wanna test

@GoogleAIStudio Its way to expensive in api even in 5.3 flash lite Way too much output tokens