Google Gemini Reduces Token Usage by Up to 88% When Analyzing Videos
**Google has incorporated agent-based video analysis into several Gemini Flash models. The tool autonomously selects relevant scenes, frames, audio, or transcripts and, according to the company, can reduce token usage by up to 88% while slightly improving accuracy in certain evaluations.
The Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite models can analyze segments of less than one second, including state changes and cuts that might be missed with conventional one-frame-per-second sampling. This capability is relevant for automated editing, where a brief transition or an instant modification can alter the meaning of a scene.
The system can also locate specific moments in hours of recording without consuming millions of tokens during the process. To do this, it identifies potentially relevant time windows and resamples them at a higher frame rate when it detects an anomaly or a signal that requires greater temporal resolution.
In addition to finding scenes, Gemini can count repeated movements and individual objects over time. The tool is not limited to observing images, as it combines frames, audio, and transcripts according to the nature of the task posed by the developer.
Google claims that this autonomous selection allows the model to focus its resources on the moments and signals that truly matter. The result is a less rigid analysis flow, where artificial intelligence decides what to examine, at what speed to do so, and through which modality before delivering a response.
Until now, Gemini used static processing that, by default, sampled video at one frame per second, although developers could adjust that setting through the API. Since the launch of native video analysis in 2025, the system combined the study of frames with the transcription of the audio track.
The agent-based approach connects the model's reasoning directly with native video tools. According to Google, Gemini can initiate a search cycle, retrieve a specific part of the file, observe the result, and decide whether it needs to consult another segment, increase the sampling speed, or change modality.
The efficiency advantages are more visible in long materials, which can range from 10-minute tutorials to 90-minute lectures and multi-hour files. With the static method, developers had to choose between incurring high token consumption or using their own selective techniques that could overlook relevant details.
Google states that its results in 1H-VideoQA and LVBench show an 88% drop in token usage, accompanied by a slight increase in accuracy. In the company's 1H-VideoQA evaluation, Gemini 3.7 Flash with agent-based analysis achieves the highest accuracy at the lowest cost per query, according to data presented by the company.
The company also included LongVideoBench among the tests used to compare the performance of the models. In that set of evaluations, Google describes Gemini 3.7 Flash as the alternative with the highest overall quality and the best combination of accuracy and cost efficiency.
These results are claims from Google and reflect its own benchmarks, so they do not independently validate all usage scenarios. Nevertheless, the potential savings can be significant for applications processing large audiovisual libraries, especially when each query requires searching for a specific moment rather than describing the entire video.
Agentic analysis is already active for uploaded videos and YouTube videos via the Gemini API in Google AI Studio and the Gemini Enterprise Agent Platform. Developers must set the processing mode to "agentic" within the API to activate selective behavior.
Google plans to extend the technology to its own products at a later date. The company expects to soon bring it to all users of the Gemini application using Flash and Flash Lite devices, while also planning to employ it in the Ask YouTube feature within the playback page in the coming months.
-- Price
This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.
You may also like

Arthur Hayes: Dollar Weakness and Yen Strength Transition Triggering Cryptocurrency Rally

Arkham API Adds Risk Score and AI Agent Payment

Hyperion DeFi Announces $20 Million Stock Buyback and Raises Full-Year Guidance

Reconstruction of the Northern Bypass of Zhytomyr: Laying of Cement Concrete Pavement

Sovereign Debt: The Secret of Gold That Governments Don't Tell

Sapien Completes New Round of Financing, Valued at $180 Million

Digital Euro Privacy: The Technical Detail That Compromises Autonomy

Etherscan Launches On-Chain Fund Tracking AI Tool Etherscan-Flow

TSLA Stock Price Prediction: Goldman Sachs vs the Analyst Who Sees Another 90% Drop

Cronos Recovers $111 Million in Stolen Funds by Rolling Back Blockchain

Police Officers Among 16 Suspects in Cryptocurrency Mining Fraud, Victims Totaling 2.7 Billion KRW

Solana Records $348 Million in RWA Net Inflows Over 30 Days, Ethereum Maintains Dominance

"Electric Vehicle Chargers as Web3 Physical Assets" ··· Park Se-jin, CEO of Gorae Protocol, "One Charger Will Become a Financial Unit"

Why Do Perpetual Contracts Experience Timed Fluctuations Every 15 Minutes?

Ethereum arbitrage study reveals builders receive $5 for every $1 burned by the network

BCRA Halts Dollar Purchases for the Third Time This Year, Ending a Streak of 27 Consecutive Positive Sessions

Stablecoin growth will test 24/7 FX liquidity, TransFi CEO says

Bitcoin Miners Invest $30 Billion in AI

BTC's 90-Day Correlation with Gold Rises to 0.56, Hitting a New High Since 2017

US Sanctions Fail to Stop $9 Billion Shadow Banking Flow Linked to Iran

Tom Lee: Ethereum is More Like Digital Land Than Digital Cash

As Polymarket and Others Step into the 'Background': The Second Half of Prediction Markets Beyond Exchanges?

Polymarket Report for the First Half of 2026: High-Frequency Traders or Super Forecasters?

Nepal Requests $2.5 Billion from UN Climate Fund After Devastating Floods

Arthur Hayes Acquires 2 Million USD in Uniswap (UNI) Over Two Days

Blockchain Capital Partner: Tokenization is the Container of Capital Markets

Bankless Founder Who Sold All ETH Turns Big Profit in Three Months with Five Assets

How to Use the WEEX Telegram Mini App: No Download Needed, Plus Rewards and an iPhone 17 Pro Chance

TokenPost Ranks Second in Global Cryptocurrency Media Traffic for August







