Gemini Omni 1.1 Flash pricing: what a second of video actually costs

A second of 720p video from Gemini Omni 1.1 Flash costs about $0.10. That’s the number to plan around, and it’s the number Google’s own pricing page quotes. There’s no free tier, so the first clip you generate while reading the docs bills to your account.

The interesting part is how Google arrives at that figure, because the billing model is token-based rather than time-based. Once you understand the mechanism, the cost levers become obvious, and one of them (drafting at 360p) cuts your iteration spend by two thirds.

The rate card

Those rates apply to gemini-omni-1.1-flash, the GA model released August 27, 2026, and to the retiring gemini-omni-flash-preview endpoint. See what shipped in the GA release for the feature side.

How video output becomes tokens

Google bills video the same way it bills text: by output tokens. A second of 720p video is priced at 5,792 output tokens. Run the multiplication:

5,792 tokens x ($17.50 / 1,000,000) = $0.1014 per second

So a 10-second 720p clip costs roughly $1.01. A 40-second scene built through four extension calls costs roughly $4.06, plus the input tokens for the prior context each turn reads.

Input is cheap by comparison. At $1.50 per million, the reference images and video you attach barely register next to the output. Don’t optimize your prompt length. Optimize how many seconds of video you generate and at what resolution.

The 360p lever

Google generates 360p previews at one third the cost of 720p and up to 60% faster. That puts a 10-second draft at roughly $0.34 instead of $1.01.

Here’s what that means in practice. Prompt iteration on video models is brutal: you rarely land the shot on the first try, and every failed attempt bills in full. A realistic session looks like twelve attempts before one keeper.

Same output, 62% less spend, and the drafts come back faster so you get through the twelve attempts sooner. Set resolution to 360p in response_format while you’re iterating and only raise it when the prompt is right. The API guide has the request shape.

The 1080p and 4K tiers are upscales of the generated frames rather than native renders, and they bill above the 720p rate. Check the current per-second figures on Google’s pricing page before you budget a 4K pipeline, since those two tiers arrived with the GA release and the published rates are the authority.

What the extension workflow costs

Scene extension is where budgets go sideways, because a 40-second video is not one request. It’s a generation plus three extension calls, each billing its own 10 seconds of output plus the prior context it reads as input.

At 720p, a full 40-second scene runs about $4.06 in output, and any attempt that drifts off-story means regenerating from whichever segment broke. Draft the whole arc at 360p first for about $1.35, confirm the story holds across all four segments, then re-render. The scene extension guide covers where extensions tend to fail.

How it compares to Veo 3.1

Both models sit behind the same Gemini API key, which makes the price gap easy to miss.

Standard Veo 3.1 costs about four times an Omni second at 720p. Veo 3.1 Lite undercuts Omni. But per-second price is not the whole comparison. Veo generates native audio and extends much further (7 seconds at a time, up to 148 seconds), while Omni gives you the conversational edit loop that Veo has no equivalent for. The full comparison works through which one to call for which job. If you already run Veo, that integration guide still applies.

Three ways teams overspend on this API

Re-rendering at full resolution to check a small change. If you’re adjusting one clause in the prompt, the 360p pass tells you whether the change worked. The 720p pass is for the version you ship.

Generating in production what you could cache. Video output is deterministic enough to reuse, and at $0.10 per second the cache pays for itself immediately. Store the generated file and key it on the prompt plus parameters. Regenerating the same clip twice is a pure loss.

Retrying on timeout without checking. Video generation is slow, and a client timeout doesn’t mean the generation failed. If your retry logic fires on timeout, you can pay twice for one clip. Raise the timeout, and check for a completed interaction before you retry.

That last one is worth catching before it reaches production. Save the request in Apidog with a realistic timeout, assert on the response shape, and run it deliberately rather than letting a retry loop discover the cost for you. Setting a spend alert in Google Cloud billing is the other half of that safety net.

Estimating a monthly bill

Work from seconds of finished video, then multiply by your draft ratio.

monthly cost = finished_seconds x $0.10
             + draft_seconds x $0.034

A team shipping 5 minutes of finished 720p video a month, with a 10:1 draft-to-final ratio at 360p:

300 finished seconds x $0.10       = $30.00
3,000 draft seconds x $0.034       = $102.00
                                     -------
                                     $132.00

Note where the money goes. Drafting dominates even at a third of the price, because you draft ten times more than you ship. If that number needs to come down, the lever is fewer attempts per keeper (better prompts, keyframe control to constrain the output), not a cheaper final render.

FAQ

Is Gemini Omni 1.1 Flash free? Not through the API. There’s no free tier for Omni, though the model is free on YouTube Shorts and YouTube Create; the free-access rundown covers those paths. Google’s text models like Gemini 3.6 Flash still offer a rate-limited free lane through AI Studio.

How much does a 10-second video cost? About $1.01 at 720p, or about $0.34 at 360p.

Why is video billed in tokens? Google prices all model output in tokens for consistency. Video output converts at 5,792 tokens per second of 720p, billed at the $17.50 per million video output rate.

Do input images and reference clips cost extra? Yes, at $1.50 per million input tokens, which is small next to video output. Reference media is not the line item to optimize.

Is 4K worth it? Only for final deliverables. 1080p and 4K are upscales, so they don’t add detail the model didn’t generate. They add file size and cost.

Does a failed generation still bill? A generation that returns video bills. Content-filter rejections and errors that produce no output shouldn’t, but watch your billing dashboard during the first week of any new pipeline rather than assuming.

The economics here reward discipline more than cleverness. Draft cheap, cache aggressively, cap your timeouts, and keep one saved request per task type so you notice the day a model update changes what you get for your $0.10. Download Apidog if you want that check running before the invoice tells you.