Research-backed calculator guide

Gemini Long Video Cost: Static vs Agentic Processing

Estimate Gemini long-video cost by separating static frame sampling from agentic navigation, tool-use tokens, model rates, and output assumptions.

What “static” and “agentic” mean for billing

Static processing is a fixed ingest pass. Gemini extracts frames at a configured rate, places them into context, and processes the submitted prompt. Google says the default is 1 FPS across models, while the API allows a custom FPS for static processing. The rate matters linearly: changing a one-hour video from 1 FPS to 0.5 FPS roughly halves the frame contribution, while changing it to 2 FPS roughly doubles it, before resolution and audio are considered. ^[3]

Resolution is another multiplier. At low media resolution, the guide documents 66 tokens per sampled frame plus 32 audio tokens per second and metadata, which it summarizes as approximately 100 tokens per second at the default 1 FPS. At high media resolution, frames use 258 tokens each and the approximate total becomes 300 tokens per second. These are planning estimates. A production estimate should use the response usage when available, especially if your request changes the default processing settings. ^[3]

Agentic processing changes the shape of the bill. Instead of loading every sampled frame into one context, the model can navigate the timeline and request only evidence relevant to the question. The usage guide describes the initial video reference and prompt as input tokens, navigation reasoning as thought tokens, and on-demand transcript, frame, and audio retrieval as tool-use tokens. The final answer remains output tokens. ^[2]

That distinction matters because “one hour of video” does not uniquely determine agentic cost. A question such as “give me a broad summary of every important scene” may require wide coverage. A question such as “when does the speaker define the new pricing tier?” may be answerable from a transcript search and a few visual windows. The same file can produce different bills under different prompts.

Google’s announcement reports up to 88% lower token consumption and up to 66% lower analysis cost on tested long-form workloads, with quality improvements in those tests. Those are maximum reported benchmark outcomes, not a per-request guarantee. ^[3] The video guide also warns that agentic navigation can increase time to first token and total token use on short clips or broad queries because of thought and tool-call round trips. ^[3]

The cost formula

For a static estimate, use this planning formula:

static input tokens ≈ duration_seconds × FPS × frame_tokens + duration_seconds × audio_tokens_per_second + metadata + prompt_tokens

Then calculate:

static cost = (static input tokens ÷ 1,000,000 × input rate) + (output tokens ÷ 1,000,000 × output rate)

At low resolution and the documented default of 1 FPS, the shortcut is approximately:

static video tokens ≈ duration_seconds × 100

At high resolution, use approximately duration_seconds × 300. The shortcut includes the documented audio and metadata approximation. If you need a more inspectable estimate, model frames and audio separately: 66 or 258 tokens per frame, plus 32 audio tokens per second, at 1 FPS. ^[3]

For agentic processing, do not pretend that duration alone is enough. Use:

agentic total tokens = initial input + thought tokens + tool-use tokens + final output

The reliable value is the actual usage payload. Before production, run representative prompts, record total_input_tokens, total_thought_tokens, total_tool_use_tokens, and total_output_tokens, and keep separate samples for short visual tasks, long targeted retrieval, and broad summaries. Google explicitly recommends inspecting interaction usage for these categories. ^[2]

The model rate also needs a date and service tier. The example below uses Google’s published global Standard rate for Gemini 3.8 Flash through December 31, 2026. The same pricing page lists different rates for other models and service tiers, and Google Cloud may show regional or platform-specific prices. Do not copy the example rate into a forecast without checking the rate card on the day you run the workload. ^{[1] [6]}

Worked cost example

Assume a 60-minute lecture, one English question, and a short text answer. The video contains speech and ordinary visual slides. The example compares a static low-resolution estimate with an agentic usage assumption. It does not include storage, upload egress, retries, application hosting, or human review.

The static result is straightforward: 360,000 video tokens plus 400 prompt tokens equals 360,400 input tokens. At $0.75 per million, input is $0.2703. The 800-token response adds $0.003, producing $0.2733.

The agentic line is deliberately labeled an assumption. The token guide gives an example in which a one-hour lecture that would use about 1.08 million tokens in static mode might use about 108,000 tokens in agentic mode, depending on the prompt and content. ^[2] The current video guide’s low-resolution shortcut gives a different planning baseline of approximately 360,000 tokens for one hour. ^[3] Those figures should not be mixed as if they were the same benchmark configuration. Here, the 108,000-token value is used only to show the arithmetic of a targeted agentic workload; the real request must replace it with observed usage.

If the agentic request instead reports 500,000 total input-side tokens after navigation, the input portion becomes $0.375 rather than $0.0813, and the total becomes $0.3780 with the same 800-token answer. That sensitivity is why a budget should store usage distributions, not only one headline percentage. A useful production worksheet records median and high-percentile agentic tokens for each workload class.

How FPS, resolution, and prompt shape change the result

FPS is a direct static cost lever. A 60-minute video at 0.5 FPS has 1,800 sampled frames; at 1 FPS it has 3,600; at 2 FPS it has 7,200. With 258 tokens per high-resolution frame, the frame component alone is approximately 464,400, 928,800, and 1,857,600 tokens respectively. Audio still adds 115,200 tokens at 32 tokens per second, and metadata and text are separate. ^[3]

That does not mean lower FPS is always cheaper in a useful sense. One frame per second can miss a brief visual event. If the question is about a tiny UI change, a fast gesture, or a momentary anomaly, the cost of a second pass or a targeted high-FPS clip may be preferable to a cheap but incomplete first pass. Agentic navigation may help by finding candidate windows and resampling them, but its extra thought and tool-use tokens need to be measured.

Prompt scope is equally important. “Summarize the entire lecture” encourages broad coverage. “Find the timestamp where the speaker explains the renewal price and quote the surrounding sentence” creates a narrower search objective. The second prompt is a better candidate for agentic routing, but it still needs verification if the result drives an editorial, compliance, or customer-facing decision.

Context windows do not erase billing. Google documents 1-million-token context windows for many Gemini models and notes that long-video handling still depends on how media becomes tokens. ^[5] A file fitting in context can still be expensive if you repeatedly resend it. For repeated questions about the same media, context caching may change the economics; Google’s long-context guidance identifies caching as a primary optimization for reused context. ^[5]

A routing worksheet for production

Start with three labels: video length, question coverage, and visual precision. Under five minutes with a broad summary or latency-sensitive response, static is a reasonable baseline. For a long lecture, interview, webinar, or meeting where the question points to one topic or moment, run agentic first and compare it with a static sample. For sub-second motion or tiny visual details, test both modes and define an acceptable review cost before choosing.

Next, capture actual usage. Do not infer agentic tokens from duration or assume the maximum headline saving. Persist the model name, service tier, processing mode, media resolution, FPS when static, prompt class, total input, thought, tool-use, output, latency, and human verification outcome. This creates a cost ledger that can be reconciled to invoices later. The AI token cost calculator is useful for separating input and output assumptions, while the AI API invoice reconciliation calculator helps compare the worksheet with provider usage records.

Finally, compare effective cost per accepted answer, not only tokens. If static costs $0.27 and agentic costs $0.08 but agentic requires a second review pass on 20% of results, the cheaper token line may not be the cheaper workflow. Conversely, a targeted agentic answer that avoids manual search across a two-hour recording may justify a higher request cost. The LLM cost calculator provides a broader monthly framing when you need to multiply the request estimate by volume.

Limitations and exclusions

This worksheet is for Gemini API token economics, not a complete video-processing total-cost model. It excludes file storage, upload and download transfer, client-side transcoding, queueing, retries, failed requests, rate-limit backoff, application hosting, observability, database indexing, persistent video search, and human review. It also assumes USD list-style rates from the cited rate card and does not model taxes, negotiated contracts, credits, or free-tier eligibility.

Pricing and supported model lists change. The values above are dated November 2, 2026 and use the stated Gemini 3.8 Flash global Standard assumption. Google’s Cloud rate card can differ by endpoint, region, model family, and service tier; its notes also distinguish long-context pricing and other platform features. ^[6] Recheck the official rate card before committing to a budget.

The static formulas are estimates, not invoice reconciliation. The agentic example is more uncertain because the model chooses what to inspect. A safe budget uses representative traces and a high-percentile allowance. Keep the original video and returned timestamps available for audit, especially when selective navigation is used to make a decision about a brief visual event.

FAQ

Is Gemini agentic video processing always cheaper than static processing?

No. Agentic processing can be substantially more token-efficient for long-form, targeted questions, but navigation adds thought and tool-use work. Google’s guide warns that agentic processing can increase total token use on short clips or broad queries. ^[3] Treat “up to” savings as benchmark context, then compare both modes on your own videos and prompts.

How do I estimate Gemini static video tokens?

Start with duration in seconds multiplied by the configured FPS and frame-token allocation. Add audio tokens and metadata. Google’s current guide gives 66 tokens per low-resolution frame or 258 tokens per higher-resolution frame, plus 32 audio tokens per second; at the default 1 FPS it summarizes those as approximately 100 or 300 total tokens per second. ^[3] Add prompt and output tokens separately when pricing the request.

What tokens are counted in agentic video usage?

The response can separate initial input tokens, navigation thought tokens, on-demand tool-use tokens, and final output tokens. Tool-use tokens can represent transcript, frames, and audio loaded during navigation. ^[2] Store those fields rather than trying to reconstruct agentic cost from video duration.

Which Gemini model rate should I use in a budget?

Use the exact model, endpoint, service tier, region, and date that your application will use. In this article’s example, Gemini 3.8 Flash global Standard is $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026 according to the cited rate card. ^[1] That assumption is not interchangeable with Gemini Pro, a non-global Cloud endpoint, Batch, Flex, Priority, or a later rate card.

Does a one-million-token context window make long video free?

No. Context capacity is not a zero-cost allowance. Video is converted into tokens, and input and output are billed according to the selected model and tier. Repeated questions can also reprocess context unless you use an appropriate caching strategy. Google’s long-context documentation presents caching as an optimization for reused context, not as a blanket waiver of input billing. ^[5]

Should I use static or agentic mode for frame-by-frame inspection?

Begin with static or a controlled custom-FPS workflow when uniform coverage is the requirement. Agentic mode is designed to navigate selectively and may not inspect every frame. For fleeting events, run a representative test, compare timestamps against the original video, and include review effort in the decision. The cheapest token result is not automatically the most reliable evidence.