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

Bitcoin ETFs: 2026 Flows Finally Return to Positive

Cosmos Hub Recovers 1.227 Million ATOM, Funds Stored in 4/6 Multisig Address

Federal Reserve Plans to Raise Regulatory Thresholds for Large Banks

INDODAX Highlights Strengthening of National Crypto Ecosystem at FEKDI x IFSE 2026 - Fintech World

Treasuries at 21-Year High: Impact on Stocks and Interest Rates

S&P 500 Closes Flat as 328 Stocks Decline

Fed proposes GENIUS Act rules for stablecoin reserves and bank issuers

xStocks adds Ledger hardware wallet support for tokenized shares

John Templeton: "Bull markets are born in pessimism"

Perpetuals on Gold, Oil, and Stocks: $117 Billion Traded in One Month

What Happens When the AI Bubble Bursts? MIT University Responds

IMF Calls for Fewer but Deeper Reforms to Address a More Vulnerable Global Economy

Meta's 'Muse' Sparks High Expectations... "An iPod Moment for AI" (Comprehensive)

大冰要抄底(专注交易) Price Prediction for October

Two men arrested for fraud with fake EURC

Mint launches connected Web3 gaming economy with MNTD rewards

Slow Fog: MemoryOS and OpenClaw Plugin Compromised

Nick Clegg's View on AI: The Real Risk is Power, Not Robots

135,694 Crypto-Millionaires, 92,272 Bitcoin-Millionaires in 2026

CFTC Chair Calls for "Mass Tokenization"! Wall Street Faces Three Barriers to Full On-Chain Adoption

AI: DeepSeek and Kimi Allegedly Redirected Queries to Claude Without Users' Knowledge

Citi Predicts Fed Will Keep Rates Unchanged in October and December, Resume Rate Cuts in June 2027

How to Rewrite Internet Rules When Everyone Has an Indefatigable Agent

Long.xyz Founder Emphasizes Asset Distribution and Scale Growth

PitchBook Predicts Kalshi Valuation Could Reach $42.1 Billion

FedNow readies cross-border support for U.S. banks







