Service eligibility and regional restrictions
PuppyIP serves only compliant overseas businesses and their authorized personnel. Proxy services are not available in mainland China. The service may only be used for lawful business activities outside mainland China. Use of this service within mainland China is prohibited.
Hosting a proxy IP or server overseas does not change these restrictions. The service must not be provided to end users in mainland China through relaying, forwarding, sharing or resale. Before use, read the Terms of Service.
Key Takeaways
- Agentic Video dynamically examines frames, audio and transcripts to locate relevant segments. Default static mode still processes at a fixed frame rate.
- Current supported models are Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite, for uploaded videos and public YouTube videos.
- Google benchmarks show up to 88% fewer tokens, 66% lower costs and about 7% higher quality for long videos. These are not guarantees for every workload.
- The feature uses standard Gemini API token pricing without a separate feature fee. Read your own usage and bills for actual costs.
- Keep a static comparison and rollback switch for short videos, full-frame audits or drifting results.
How Agentic differs from Static
Static is the default: it samples at a fixed frame rate and places content into context at once. Agentic dynamically searches the timeline based on the question, selecting transcripts, audio or frames and adjusting viewing speed, frame rate and resolution. It better suits finding moments, detecting anomalies, counting and searching across hours of video.
A pipeline that sparsely samples frames to save tokens may miss brief actions, while higher FPS increases costs. Agentic lets the model decide where to look based on the question. Start by establishing an A/B baseline with the same video and question, rather than deleting the old pipeline.
Supported scope and activation
Officially supported models are Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite. In Gemini API Interactions calls, set the video input object's processing to agentic. The announcement covers uploaded videos and public YouTube videos. Do not assume similarly named models support it.
Gemini App rollout is separate from API availability. The announcement says the app will roll out gradually. An absent entry point for one account is not necessarily an API fault, and does not justify claiming simultaneous access across all regions, plans and clients.
Which tasks should try Agentic first
Prioritize videos longer than ten minutes, brief state-change detection, audio/frame cross-checks, or tasks whose static requests already consume many tokens. Examples include course highlights, meeting navigation, anomalies in long recordings, sports-action counts and candidate editing points.
Keep static first for global descriptions of short clips, regulatory full-frame coverage, reproducible fixed sampling or custom-FPS dependencies. Agentic's dynamic choices change the evidence the model sees, so cost reduction cannot be the only acceptance metric.
Validating tokens, cost and quality
Build at least three sample groups: long-video questions, moment localization and counting. Hold video, prompt, model and scoring constant, run static and agentic, and record total tokens, latency, usage cost, referenced timestamps, misses and false positives. Google's 88%, 66% and approximately 7% figures are benchmark maxima, not promised outcomes for your material.
Standard token pricing without an extra feature fee does not make requests free. Long-video uploads, repeated replays, output length and failed retries still generate usage. Bills, usage fields and the current pricing page provide evidence of your own costs.
Seven checks before release
In order: confirm a supported model; save the static baseline; select an authorized video source; enable agentic for one request; compare answers and usage; add timeout, retry and content-permission checks; then stage rollout only for suitable tasks. Do not change processing across every video task at once.
Record video URI, model version, request ID, processing mode, question, returned timestamps and human acceptance results. For YouTube URLs, investigate public-video accessibility separately from API identity permissions.
Separating errors and anomalous results by layer
Unsupported parameters or models usually concern request configuration. Unreadable video may involve URI, upload status, permissions or format. 429 and 503 concern quota or service layers. Inspect the network only with evidence of connection timeouts, DNS or TLS issues. Save the complete error body and request ID rather than masking schema problems by switching exits.
Use the Gemini 429/503 troubleshooting guide to separate quota and service errors, then the AI API configuration guide to check endpoint, key and model.
Stop and rollback conditions
Stop expansion and return to static when important moments are missed, referenced timestamps cannot be reproduced, costs rise, latency exceeds limits, answers vary beyond tolerance, or audits cannot accept dynamic sampling. Control rollback by task type instead of hastily changing every prompt.
For stable access to Google AI Studio or overseas developer documentation, teams can visit the PuppyIP website to learn about fixed network exits. The network can improve connection paths but cannot change model support, quotas, content permissions or Agentic processing semantics.
Sources
Frequently Asked Questions
Which models support Gemini Agentic Video?
Google currently lists Gemini 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite. Check current official documentation for other models.
Does omitting processing automatically use Agentic?
No. Video defaults to static processing. Explicitly set processing to agentic in a supported model's video input.
Will Agentic Video always save 88% of tokens?
No. 88% is Google's highest result on particular long-video benchmarks. Retest with your own videos, questions and usage data.
Does Agentic Video have an extra feature fee?
Google says standard Gemini API token pricing applies without a separate feature fee. Still account for request tokens, uploads and retries.
Can it use YouTube videos?
The announcement includes public YouTube and uploaded videos. Verify URL accessibility, content permissions and current API support.
When should we return to static?
Return to the verified static path for missed key frames, irreproducible results, cost or latency overruns, or dynamic sampling that fails audit requirements.