Google DeepMind Launches Agentic Video for Gemini: 88% Fewer Tokens, 66% Lower Cost

On September 1–2, Google DeepMind introduced Agentic Video for its latest Gemini models and the Gemini API. Instead of statically processing an entire video, the model actively inspects, locates, and reasons over temporal visual context—deciding what to watch, at what speed (adaptive frame rate), and with which signals (transcripts, audio, frames)—maintaining SOTA video understanding while dramatically cutting cost. Developers can enable it per video by setting Video processing to "agentic"; Google shipped a blog post, a developer guide, and AI Studio support alongside.

Confirmed

• Official figures: video analysis cost down up to 66%, token usage down up to 88%, accuracy up up to 7% (relayed by @rseroter, @patloeber, @MarioLucic, @SabooShubham, @mfckreth, @CSProfKGD, among others).
• @OfficialLoganK said the feature is available in the Gemini API as "Agentic Video," controllable per video, supported on 3.7 Flash and other models; Google published an official blog post the same day.
• @philschmid's breakdown: the model iteratively navigates the timeline, choosing what to view, frame rate (0.1 or 10 FPS), and whether it needs speech transcription, audio, or visual frames—yielding up to 88% token savings, 66% cost reduction, and 7% accuracy gain on long videos.
• @osanseviero: rather than a fixed 200K tokens per long video, the model intelligently invokes tools (transcript analysis, adaptive FPS) based on the prompt, cutting tokens by 88% and improving latency; a follow-up noted the developer guide is out.
• Per @gaganghotra, the rollout covers Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite, replacing fixed 1 FPS sampling. @Servola-Journal added the capability shipped to those three models on Sept 1, before the much-discussed 3.8 Flash on Sept 2.
• On Sept 3, @shaneguML relayed a Google AI Developers demo: for transient actions like claps easily missed or confused at fixed 1 FPS, Gemini 3.7 Flash auto-adjusts playback speed and counts every clap precisely.
• @rseroter and @AIAndrew relayed a demo by Googler vamsibatchuk: a retro Windows 98–style workstation combining Gemini 3.8 Flash with Agentic Video Understanding autonomously decomposes a two-hour football match and pinpoints every yellow card at sub-second precision.
• @patloeber shared Maarten Grootendorst's explainer video; Reddit user @otarU posted a Gemini 3.7 Flash demo.

Unconfirmed

• Posts differ on supported models: "Gemini 1.5 Flash / 1.5 Flash-8B / 1.5 Flash-Lite" (@Scobleizer relay) versus "Gemini 3.7 Flash / 3.6 Flash / 3.5 Flash-Lite" (@gaganghotra, @SabooShubham, @OfficialLoganK, @Servola-Journal). Official launch timing and model coverage should follow @GoogleDeepMind's announcements.

Why it matters

• Video is the most token-hungry modality in multimodal apps; near-90% token savings dramatically improve the feasibility and economics of long-video agents and video-native AI products, and the API-level switch plus guides lower the barrier for developers.

2026-09-02 ~ 2026-09-03 · 31 related posts

• Google releases Gemini 3.7 Flash with cheaper, more accurate video understanding — Saboo_Shubham_ · 2026-09-02
• Google introduces Agentic Video in Gemini 3.7 Flash: fewer tokens, higher accuracy — Saboo_Shubham_ · 2026-09-02
• Gemini Agentic Video Cuts Analysis Costs by 66%, Token Use by 88% — rseroter · 2026-09-02
• [source] Gemini Adds Agentic Video Understanding, Cuts Token Usage by 88% — GoogleDeepMind · 2026-09-02
• [source] Gemini agentic video understanding: 88% fewer tokens, 66% lower cost, ~7% higher accuracy — _philschmid · 2026-09-02
• Google Launches Agentic Video Understanding for Gemini — Scobleizer · 2026-09-02
• Gemini adds agentic video understanding, cutting token usage by 88% — osanseviero · 2026-09-02
• Gemini agentic video understanding launches with a developer guide — osanseviero · 2026-09-02
• [source] Gemini API launches Agentic Video, cutting token usage by 88% — OfficialLoganK · 2026-09-02
• Google blog details Agentic Video tech for long-form processing — OfficialLoganK · 2026-09-02
• Gemini Adds Agentic Video Understanding, Cuts Tokens by 88% — AI_Andrew · 2026-09-02
• Google launches Agentic Video in Gemini: up to 88% fewer tokens, 66% lower cost — gaganghotra_ · 2026-09-02
• Google introduces agentic video understanding with Gemini — leslysandra · 2026-09-02
• Gemini API adds Agentic Video, cuts token use by 88% — RubenEVillegas · 2026-09-02
• Gemini 3.7 Flash Demonstrates Agentic Video Understanding — otarU · 2026-09-02
• Gemini launches new video tool detecting fast movements — jocarrasqueira · 2026-09-02
• Google ships agentic video understanding with Gemini, explainer video released — patloeber · 2026-09-02
• Gemini learns to 'watch' video selectively: 88% fewer tokens, 66% lower cost — mfckr_eth · 2026-09-02
• DeepMind adds agentic video understanding to Gemini: 88% fewer tokens, 66% cheaper — petrusenko_max · 2026-09-03
• Gemini Flash models now agentic video understanding: model picks frames, audio or transcript — technextpreneur · 2026-09-03
• Google demos Gemini agentic video understanding that auto-adjusts speed to count claps — shaneguML · 2026-09-03
• Google's Active Video Understanding Cuts Gemini Token Use by 88% While Boosting Accuracy 7% — APPSO · 2026-09-03
• Gemini's 88% video token cut landed on 3.7 Flash, not the 3.8 everyone's talking about — Servola-Journal · 2026-09-03
• Gemini's agentic video understanding dissects 2-hour soccer matches automatically — AI_Andrew · 2026-09-03

7 near-duplicate retellings: haider1 · patloeber · gaganghotra_ · Scobleizer · MarioLucic_ · CSProfKGD · rseroter