Gemini Omni 1.2 Flash: Pricing, Limits and How to Build With It

Gemini Omni 1.2 Flash is Google DeepMind's new video generation and editing model. It became generally available on the Gemini API and Google Cloud on November 5, 2026. It makes clips of up to 15 seconds with synchronized audio, and 720p output costs about $0.07 per second, roughly 30% less than Omni 1.1 Flash. For builders, the main points are the per-clip cost, a different API from the rest of Gemini, several region and feature limits, and a forced migration from 1.1 on December 3, 2026.

The launch did well on Hacker News, but most coverage so far repeats Google's announcement. This guide covers what you can actually build, what it costs per clip, and what to check before you ship.

What Google shipped

According to Google's announcement, Omni 1.2 Flash generates and edits video natively up to 15 seconds long, up from 10 seconds in Omni 1.1. You can set the length in one-second steps from 3 to 15 seconds. The main changes:

• Reference adherence. You can supply character, product or style images, plus short motion clips. Google says faces, garments, packaging and brand colors stay consistent across shots.
• Less drift. Google says there is less morphing from the first frame to the last, including when you extend a clip or interpolate between keyframes.
• Multi-turn editing. You refine a clip turn by turn instead of regenerating it from scratch, for example swapping a product, changing the background or relighting the scene.
• Speech. Google promises more natural dialogue, clearer lip-sync and conditioning for multiple speakers.
• Draft, then upscale. You draft at 360p or 720p and upscale only the keeper to 1080p or 4K.

Every output carries an imperceptible SynthID watermark. Google Cloud's model card also lists C2PA Content Credentials as supported.

The model is rolling out in several places. Developers get the Gemini API, Google AI Studio and the Google Cloud console. Google's own apps include Flow, Google Vids (15-second clips at no extra cost with a Google account or Workspace subscription), the Gemini app for Google AI Ultra subscribers, and Google Ads asset studio.

Treat the quality claims as unverified for now. Google has published no formal benchmarks, and The Decoder points out that the claims are Google's own, with no independent testing yet. The Google Cloud launch post quotes early customers including Photoroom, Picsart, invideo, Krea, FLORA, Artlist and QuickFrame AI, and names Runway as a user. Those are endorsements, not evaluations. Google also says it is still working on fast sports action, rapid script changes and complex physics.

What it really costs per clip

The headline "$0.07 per second" is only the 720p rate. The Gemini API pricing page bills video by output tokens: $12.00 per million video output tokens and $1.50 per million input tokens (text, image or video). There is no free tier for this model. Each resolution uses a fixed number of tokens per second, so the per-clip math is predictable:

For comparison, the same page lists Omni 1.1 Flash video output at $17.50 per million tokens, about $0.10 per second at 720p. The drop to $12.00 is where the "30% cheaper" claim comes from. Input tokens add a little on top, which matters most when each edit turn sends video back in.

Why the draft-then-upscale workflow matters for your margins

At 4K, a 15-second clip costs about nine times as much as a 360p draft. If your product lets users iterate ("make the lighting warmer", "swap the sneaker"), do every iteration at 360p and render the final clip at a high resolution only once.

Here is an illustrative budget for one finished 15-second ad. It assumes the final upscale bills like a full 1080p generation, and Google's pricing page doesn't break out upscaling separately, so check this on your own bill:

1. Four 360p drafts to choose a direction: about $1.39
2. Two 360p edit turns on the chosen draft: about $0.70
3. One upscale to 1080p: about $1.56

That totals roughly $3.65, compared with about $9.40 for doing all seven steps at 1080p. If you charge per video or per credit, plug these numbers into a cost model before you set prices. The LLM API Cost Calculator is useful for the rest of your AI spend. For video, multiply seconds by the token rate in the table above.

Developer details that break existing code

This is not a drop-in Gemini model. Google Cloud's model card lists only the Interactions API and Count Tokens API as supported. The generateContent API, the Live API, the OpenAI-compatible Chat Completions endpoint and batch inference are all unsupported. The model ID is gemini-omni-1.2-flash.

The model card also says it does not support system instructions, structured output, context caching, function calling, grounding or tuning. The Gemini API Omni docs add that temperature, top_p, stop sequences and negative prompts are unsupported too. Write your exclusions into the prompt itself, for example "No dialogue."

Hard limits to design around

According to the Gemini API Omni docs:

• Duration: 3 to 15 seconds in whole seconds, defaulting to 10. First/last-frame interpolation and each extension turn are limited to 3 to 10 seconds. Asking for more returns an invalid-argument error.
• Reference images: up to 6 in total, with at most 3 per character and at most 2 characters.
• Reference videos: up to 3 clips of 3 seconds each for outputs of 3 to 10 seconds, but only 1 for outputs of 11 to 15 seconds. Reference videos work only for new generations, not for edits or extensions.
• Extension: uses the last 10 seconds of the input, adds 3 to 10 seconds per turn and tops out at 30 seconds in total. You can only append to the end of a clip, not the start or the middle.
• Aspect ratios: 16:9 and 9:16 only, according to the Cloud model card.
• Language: only English is fully supported. Other languages haven't been evaluated.

Multi-turn editing and storage gotchas

To edit a generated clip, pass previous_interaction_id with your new prompt. If you set store=false on the original call, you can't edit that clip later. That is a real trade-off if you would otherwise turn off storage for privacy reasons.

Delivery is inline base64 by default. For files over 4MB, which covers anything above 720p, request "delivery": "uri". A later GET on the interaction returns inline data even if you asked for a URI, so save the URI from the creation response. The docs say it's only guaranteed there or in the SSE stream.

Prompting changes

Google now says the angle-bracket control tags, such as FIRST_FRAME, IMAGE_REF_0 and VIDEO_REF_0, are deprecated and won't be supported in future models. Describe each input's role in plain language instead, such as "use the first image as the opening frame." For edits, the docs suggest short prompts and adding "Keep everything else the same."

Where the sources disagree

Google's marketing pages, the Gemini API docs and the Cloud docs don't fully agree. Check the surface you're actually building on:

In practice, the extra capacity guarantees and gated features sit on Google Cloud. Production workloads that need reserved throughput or SLAs, a point Photoroom's CEO raises in the Cloud blog, should probably run there rather than on an AI Studio key.

Region and compliance constraints

Two limits will affect some products directly:

• EEA, Switzerland and UK: according to the Gemini API docs, users there can't upload a video and then edit or extend it, and they can't upload or edit images containing minors. Extending videos the model itself generated works in all available regions. If your product is "upload your clip and let AI edit it", you need region detection and a fallback flow for those users.
• Global endpoint only on Google Cloud: the model card lists global as the only region, for both Provisioned Throughput and pay-as-you-go. You can't pin processing to an EU or US region, which can matter for enterprise customers with data residency requirements.

On provenance, SynthID watermarking and C2PA support help if your customers run ads on platforms that require AI-content disclosure. They don't replace your own labeling.

What to do now

1. Migrate before December 3, 2026. The pricing page says that from that date, all API requests to gemini-omni-1.1-flash will route to 1.2 Flash and bill at 1.2 rates. Your price drops, but the outputs will change. Run your saved test prompts against 1.2 now, strip out deprecated control tags, and check how faces and products look in your typical shots.
2. Rebuild your cost model around tokens per second. Default user drafts to 360p, put 1080p and 4K behind an explicit "finalize" step, and log output tokens per job so you can confirm how upscaling actually bills.
3. Plan your storage. Keep store enabled for anything a user might edit, save URIs when you create them, and decide how long you keep generated media.
4. Decide between the Gemini API and Google Cloud early. Throughput guarantees, allowlisted audio references and enterprise controls sit on the Cloud side. A quick AI Studio prototype may not match production behavior.
5. Compare before committing. Without published benchmarks, run your own blind comparison against the video models you already use, scoring on the shots your customers actually ask for. You can browse alternatives in the AI video tools directory.

Over the coming weeks, watch for independent quality tests, for whether Google publishes separate pricing for upscaling and extension, and for whether the Gemini API catches up with Cloud on Provisioned Throughput and audio references. Those three things will decide whether Omni 1.2 Flash is a cheap drafting engine or your main production model.