It's worth naming the actual model identifiers rather than speaking vaguely about "the Omni model," since Google itself uses several names across its own documentation. The current, generally available model on the Gemini API is gemini-omni-1.1-flash. The original preview identifier, gemini-omni-flash-preview, still appears in Google Cloud's own model card for the Gemini Enterprise Agent Platform, where it's explicitly labelled a Generative AI Preview offering with its own separate terms.

On that enterprise platform specifically, the model card lists a 131,072-token maximum input and a roughly 57,920-token maximum output — real technical ceilings worth knowing if you're planning a production integration rather than experimenting casually.

Here's what's worth being honest about rather than guessing at: Google hasn't published a technical paper detailing Omni's exact underlying architecture, the way it has for some other model releases. What's confirmed is that it sits within the Gemini 3.x-era model family by both timing and Google's own positioning, sitting in Google's model documentation alongside Gemini 3.5 through 3.8 Flash variants. Beyond that, claims about its precise internal design would be speculation, and this article isn't going to pretend otherwise.

There are three genuinely separate on-ramps here, serving different audiences, and it's worth being clear about which one actually fits your situation rather than defaulting to whichever is easiest to find.

For everyday use, the Gemini app and Google Flow are the two consumer paths, both included in a Google AI Plus, Pro, or Ultra subscription, alongside a completely free path through YouTube Shorts and the YouTube Create app — no subscription required for that specific route.

For developers, an API key from Google AI Studio and a short script against the Interactions API is the fastest path to testing it directly:

from google import genai

client = genai.Client()

A single text-to-video call against the current GA model

interaction = client.interactions.create( model="gemini-omni-1.1-flash", input="A drone shot of a mountain landscape at sunrise." )

with open("sunrise.mp4", "wb") as f: f.write(interaction.output_video.data)

client.interactions.create is the core call, and the model string is the one line that matters most to get right, since gemini-omni-1.1-flash is the current general availability identifier, while some existing tutorials and cached documentation still reference the older gemini-omni-flash-preview string from the June preview window. The response comes back as raw video bytes accessible through interaction.output_video.data, written directly to a file here, though for anything beyond a quick test, Google's documentation recommends URI-based delivery instead of inline bytes, to avoid hitting payload size limits on larger outputs.

For enterprise teams already operating inside Google Cloud's governance and billing structure, the Gemini Enterprise Agent Platform — the successor branding for what was Vertex AI — is the intended path, offering the same model behind Google Cloud's existing access control and monitoring tooling.

Here's the actual rate card rather than a vague "it's affordable" claim. On the Gemini API's paid tier, input is billed at $1.50 per million tokens, text and thinking output at $9 per million tokens, and video output at $17.50 per million tokens. Video is billed by a fixed conversion rate of 5,792 tokens per second of 720p output, which works out to roughly $0.10 per second — a figure Google states directly on its own pricing documentation.

Worked out concretely, a 10-second 720p clip costs approximately $1.01 in video output alone, before adding whatever small amount the input prompt and any reference images contribute at the much lower input rate. Unlike most other models in the Gemini lineup, Omni currently has no free tier on the API, so the very first test call a developer makes bills to their account, not a rate-limited trial allowance. The practical cost lever worth knowing is the 360p draft mode introduced in the August general availability update — meaningfully cheaper per second than 720p, and useful for iterating on a prompt before committing to a full-resolution render of the version you actually want to keep.