--- title: Gemini now skims videos like a developer skims docs, and... description: Gemini's video understanding now supports an 'agentic' processing mode that mimics how a person skims a video: scanning the transcript, jumping to relevant... canonical: https://daily.dev/posts/gemini-now-skims-videos-like-a-developer-skims-docs-and-the-token-savings-are-real-tbebtdwac twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Gemini now skims videos like a developer skims docs, and the token savings are real | daily.dev og:description: Gemini's video understanding now supports an 'agentic' processing mode that mimics how a person skims a video: scanning the transcript, jumping to relevant... og:url: https://daily.dev/posts/gemini-now-skims-videos-like-a-developer-skims-docs-and-the-token-savings-are-real-tbebtdwac og:image: https://api.daily.dev/og/posts/tbebTDwAC.png og:image:alt: Gemini now skims videos like a developer skims docs, and the token savings are real og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Gemini now skims videos like a developer skims docs, and the token savings are real **[Trends](https://daily.dev/sources/trends)** · 2 min read · 0 upvotes · 0 comments ## Summary Gemini's video understanding now supports an 'agentic' processing mode that mimics how a person skims a video: scanning the transcript, jumping to relevant timestamps, and pulling frames or audio only when needed via a think-act-observe loop with tools like get_transcript, get_frames, and get_audio. It even self-selects frame rate (0.1 to 10 FPS). Reported results: 88% fewer tokens, 66% lower cost, and roughly 7% higher benchmark accuracy versus the old brute-force frame-feeding approach. Enable it by setting processing="agentic" on the video input; Google recommends static mode for clips under two minutes. Available now in the Gemini API and Google AI Studio for Gemini 2.5 Flash, 2.0 Flash, and 1.5 Flash Lite. ## Content Google just shipped agentic video understanding for Gemini, and the numbers are the kind that make you do a double-take: 88% fewer tokens, 66% lower costs, and roughly 7% better accuracy on benchmarks. The trick is almost embarrassingly obvious in hindsight. Instead of dumping an entire video into the model as a wall of frames, Gemini now runs a think-act-observe loop where it decides what to actually look at. It can call `get_transcript` to scan speech first, pinpoint the relevant timestamps, then fetch only those frames at whatever frame rate makes sense (0.1 FPS for slow-moving content, 10 FPS when motion matters). If acoustic cues are what the query needs, it pulls audio directly. The model navigates the video the way a human researcher would skim a lecture recording, not the way a scanner processes a document. Phil Schmid's summary on X captures the practical upside well: the model "scans speech transcripts to pinpoint relevant moments before fetching visual frames" rather than paying the full token cost upfront. For a two-hour lecture or a multi-hour security recording, that's the difference between a query that's economically viable and one that isn't. Enabling it is a single parameter: set `processing="agentic"` on the `video` object. Google recommends keeping the default `static` mode for anything under two minutes, where the overhead of the agentic loop isn't worth it. The feature is live now in the Gemini API and Google AI Studio across Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash Lite, at standard token pricing with no extra fee. Longer-term, it's slated to power YouTube's "Ask YouTube" feature and roll into the Gemini app. The use cases Google is highlighting - needle-in-a-haystack search across long recordings, anomaly detection, object counting - are exactly the ones where static frame processing was previously either too expensive or too slow to be practical. Whether the 7% accuracy gain holds outside Google's own benchmarks is the question worth watching, but the cost reduction alone is enough to make previously unworkable video pipelines worth revisiting. ## Questions this post answers ### How do I enable agentic video processing in the Gemini API? Set processing="agentic" on the video object when making a request. This enables a think-act-observe loop where Gemini decides what to examine rather than processing every frame, first scanning the transcript via get_transcript to find relevant timestamps, then fetching only those frames at a suitable frame rate (as low as 0.1 FPS or as high as 10 FPS), or pulling audio when needed. The feature is live in Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash Lite at standard token pricing. _Developers optimizing video-analysis pipelines can track Gemini API changes like this via daily.dev._ ### How much does agentic video processing reduce Gemini API costs compared to static frame processing? It reduces token usage by 88% and cost by 66%, while also improving accuracy by roughly 7% on Google's benchmarks. The savings come from having the model selectively fetch only relevant frames and transcript segments instead of processing an entire video as a wall of frames, which matters most for long content like two-hour lectures or multi-hour security recordings. _Anyone weighing video AI pipeline costs can follow benchmarks like these on daily.dev before committing._ ### Should I use agentic or static video processing mode in the Gemini API for short clips? Use the default static mode for videos under two minutes, since Google states the overhead of the agentic think-act-observe loop isn't worth it at that length. Agentic mode is intended for longer content, such as multi-hour recordings, where selectively fetching frames and transcripts saves significant tokens and cost versus processing every frame upfront. _Teams choosing between processing modes can compare tradeoffs like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 2 discussions and 18 comments across x (as of 2026-09-13). **TL;DR:** Reactions are mostly curious and mildly enthusiastic about the token/cost savings, with a few people questioning whether this alone shifts model preference or asking for missing features like local models and video editing. **Sentiment:** 45% positive · 45% mixed · 10% skeptical **The case for** - Some see the fps/transcript-first approach as a genuinely useful way to cut token usage and latency. - A few note the accuracy and cost improvements are impressive for video processing workflows. **The pushback** - One person argues the update won't be enough to pull developers away from OpenAI and Anthropic. - Another wonders how much of the improvement is from skipping boring parts versus smarter frame-rate selection, suggesting the mechanism isn't fully clear. - Requests for an open-weight or local model and video editing capability suggest the feature is seen as incomplete. **By community** - x (mixed): Curiosity about YouTube support and mechanism mixes with genuine enthusiasm over token savings and some skepticism that this changes developer platform choice. - x (mixed): Praise for the cost/accuracy gains is tempered by a pointed comment that this won't win over people already committed to other providers, plus requests for open-weight or local options. **Hottest debate:** Whether meaningful token/cost savings on video processing are enough to change which AI provider developers actually build with. **Open questions** - How much of the savings comes from skipping irrelevant footage versus simply choosing a smarter frame rate? - Will an open-weight or local/on-device version of this capability be offered? - Can the feature support video editing, not just analysis? **Highlights** > @_philschmid Cool update, but I don't think people will be switching for this. OpenAI and Anthropic are still the ones people actually build with. Gemini getting cheaper on video doesn’t change that. Google comeback soon tho..? > — [minkrov on x · 1 points](https://x.com/minkrov/status/2094857621031289239) > @osanseviero curious how much of the win is "don’t watch the boring parts" vs smarter fps 🤔 > — [ok1mraise on x](https://x.com/ok1mraise/status/2094844742911533184) > @_philschmid Will there be an open-weight model one day? > — [NotPhilSledge on x · 1 comments](https://x.com/NotPhilSledge/status/2094865065069461822) > @osanseviero Can it edit videos or not yet? > — [Detffhted on x](https://x.com/Detffhted/status/2094876477326041285) > @osanseviero 88% fewer tokens by letting the model pick fps and transcription instead of dumping the whole video. that's the version of agentic that actually saves money > — [Mirabuildz on x](https://x.com/Mirabuildz/status/2094868222453387321) **Source threads** - [x](https://x.com/osanseviero/status/2094841663399022757) · 0 points · 9 comments - [x](https://x.com/_philschmid/status/2094840645617619036) · 0 points · 9 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#google-gemini](https://daily.dev/tags/google-gemini), [#google-ai-studio](https://daily.dev/tags/google-ai-studio) [View this post on daily.dev](https://daily.dev/posts/gemini-now-skims-videos-like-a-developer-skims-docs-and-the-token-savings-are-real-tbebtdwac) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Gemini now skims videos like a developer skims docs, and the token savings are real","url":"https://daily.dev/posts/gemini-now-skims-videos-like-a-developer-skims-docs-and-the-token-savings-are-real-tbebtdwac","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/gemini-now-skims-videos-like-a-developer-skims-docs-and-the-token-savings-are-real-tbebtdwac"},"datePublished":"2026-09-01T17:36:14.882Z","dateModified":"2026-09-13T20:04:29.931Z","description":"Gemini's video understanding now supports an 'agentic' processing mode that mimics how a person skims a video: scanning the transcript, jumping to relevant...","image":"https://pbs.twimg.com/media/HRJW8RpWIAAJq00.jpg","thumbnailUrl":"https://pbs.twimg.com/media/HRJW8RpWIAAJq00.jpg","isAccessibleForFree":true,"articleSection":"Trends","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/gemini-now-skims-videos-like-a-developer-skims-docs-and-the-token-savings-are-real-tbebtdwac","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,google-gemini,google-ai-studio","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"Gemini now skims videos like a developer skims docs, and the token savings are real"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/gemini-now-skims-videos-like-a-developer-skims-docs-and-the-token-savings-are-real-tbebtdwac#faq","mainEntity":[{"@type":"Question","name":"How do I enable agentic video processing in the Gemini API?","acceptedAnswer":{"@type":"Answer","text":"Set processing=\"agentic\" on the video object when making a request. This enables a think-act-observe loop where Gemini decides what to examine rather than processing every frame, first scanning the transcript via get_transcript to find relevant timestamps, then fetching only those frames at a suitable frame rate (as low as 0.1 FPS or as high as 10 FPS), or pulling audio when needed. The feature is live in Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash Lite at standard token pricing. Developers optimizing video-analysis pipelines can track Gemini API changes like this via daily.dev."}},{"@type":"Question","name":"How much does agentic video processing reduce Gemini API costs compared to static frame processing?","acceptedAnswer":{"@type":"Answer","text":"It reduces token usage by 88% and cost by 66%, while also improving accuracy by roughly 7% on Google's benchmarks. The savings come from having the model selectively fetch only relevant frames and transcript segments instead of processing an entire video as a wall of frames, which matters most for long content like two-hour lectures or multi-hour security recordings. Anyone weighing video AI pipeline costs can follow benchmarks like these on daily.dev before committing."}},{"@type":"Question","name":"Should I use agentic or static video processing mode in the Gemini API for short clips?","acceptedAnswer":{"@type":"Answer","text":"Use the default static mode for videos under two minutes, since Google states the overhead of the agentic think-act-observe loop isn't worth it at that length. Agentic mode is intended for longer content, such as multi-hour recordings, where selectively fetching frames and transcripts saves significant tokens and cost versus processing every frame upfront. Teams choosing between processing modes can compare tradeoffs like this on daily.dev."}}]} ```