Google DeepMind Launches Agentic Video Understanding for Gemini, Cutting Token Use by Up to 88%
On September 2, Google DeepMind announced "Agentic Video Understanding" for the Gemini model family, letting the model dynamically decide which parts of a video to watch, at what speed, and with which signals (speech, audio, frame content) to analyze—boosting accuracy while sharply cutting the token consumption and cost of video analysis.
Confirmed
- Official figures show the feature can cut video analysis costs by up to 66%, reduce token consumption by up to 88%, and improve accuracy by up to 7%.
- According to @Scobleizer, the feature ships on models such as Gemini 1.5 Flash, 1.5 Flash-8B and 1.5 Flash-Lite.
- @SabooShubham demonstrated the capability using Gemini 3.7 Flash, reporting major drops in cost and token consumption for video understanding tasks along with better accuracy; @rseroter also relayed the official cost and accuracy figures.
Unconfirmed
- Posts mention both "Gemini 1.5 series" and "Gemini 3.7 Flash" model names; the official launch scope and applicable model versions should be confirmed against @GoogleDeepMind's announcement.
Why it matters
- Video is one of the most token-hungry modalities in multimodal applications; nearly 90% token savings significantly improves the viability and economics of video-understanding agents and long-video analysis, potentially enabling more video-native AI products.
2026-09-02 ~ 2026-09-02 · 14 related posts
Primary sources
- Gemini Adds Agentic Video Understanding, Cuts Token Usage by 88% — GoogleDeepMind ·
- Gemini API launches Agentic Video, cutting token usage by 88% — OfficialLoganK ·
- Gemini adds agentic video understanding, cutting token usage by 88% — osanseviero ·
- 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
- Google Launches Agentic Video Understanding for Gemini — Scobleizer · 2026-09-02
- [source] 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
3 near-duplicate retellings: haider1 · patloeber · gaganghotra_