← Back to blog
NewsAbout 7 min read

Gemini Omni 1.2 Flash Killed the Video Slot Machine

Published Nov 7, 2026
Gemini Omni 1.2 Flash Killed the Video Slot Machine

Google made Gemini Omni 1.2 Flash generally available on November 5, and the pricing structure matters more than the model.

The headline specs are straightforward. Native clips now run to 15 seconds, up from 10. Duration can be set anywhere between 3 and 15 seconds in one-second steps, with 10 seconds as the default. Output defaults to 720p at 16:9, with 9:16 supported. You can draft at 360p or 720p and upscale selected results to 1080p or 4K. All output carries an invisible SynthID watermark.

Then the number: 720p generation at $0.07 per second, and 360p at about $0.023. On the paid API, that is $12 per million video output tokens, metered at 5,792 tokens per second of 720p, which works out to roughly $0.70 for a ten-second clip and $1.04 for fifteen. Google says it is 30% cheaper than Omni 1.1 Flash.

Compare that with where the market was a year ago. A single retry on a flagship video model cost enough that creators learned to write prompts defensively and accept the third-best result. The industry called it 抽卡, drawing cards, because that is what it was.

The retry becomes affordable, so retries change

At seven cents a second, a 360p five-second draft costs about twelve cents. Fifteen drafts cost less than two dollars. The rational workflow stops being "write one perfect prompt and hope" and becomes "generate twelve, look at them, upscale the one that works."

Six blank translucent film frames in a row on a steel table, the last one lit from behind.

That is a workflow change disguised as a price change. It moves spend from generation to selection, and selection is where human taste actually adds value. It also changes what tools get built: a browser that shows you twelve cheap drafts and one expensive final render is a different product from a chat box that charges you for every attempt.

Google is explicit about this framing. The model offers a graded workflow, and higher-resolution rendering is meant to be reserved for outputs you have already chosen.

15 seconds changes the unit of work

The jump from 10 to 15 seconds sounds incremental. It is not, because of what fits inside it.

A ten-second clip can demonstrate a product or establish a mood. It cannot comfortably introduce, demonstrate and close. Fifteen seconds is roughly the length of a standard social ad, which means a single generation can now produce the deliverable rather than a component of it. That removes the join, and joins are where AI video falls apart: lighting shifts, faces drift, objects move between cuts.

Google's own example is a 15-second virtual clothing try-on built from one image of a person and three images of outfits, with the person's identity and the garment details holding across the movement.

The consistency claims are specific. Google's side-by-side evaluations give Omni 1.2 a 62% visual-consistency win rate against 38% for 1.1 in text-to-video testing. The reference-to-video comparison gives 1.2 a 63% win rate on multi-reference identity consistency against 37%. Photoroom's chief executive Matt Rouif told reporters that longer generation was the single most requested feature from users, and that general availability with service-level agreements is what lets his company put the model in production for sellers.

Arena agrees, which is unusual

Arena.ai has Gemini Omni 1.2 Flash first in text-to-video with 1,541 points, up from 1,514 for 1.1, which was second. It also ranks first in image-to-video and third in video editing. With this release, Google's Omni family holds the top three places in text-to-video.

That is a notable reversal. Earlier this year the narrative was that Chinese labs had taken the video lead on blind preference. On Arena's numbers as of this release, that is no longer true at the top of the text-to-video board.

Leaderboards measure short clips against a prompt, which is not the same as holding a character steady across forty shots. Treat the ranking as a signal about aesthetic preference, not about production reliability.

Editing is where the practical gain sits

The feature most likely to change daily work is conversational editing. Instead of generating a new clip when something is wrong, you continue working on the existing one: replace a product, swap the background, change lighting or weather, switch aspect ratio, adjust camera angle. Unmentioned elements are supposed to survive the edit.

Technically this runs through a `previous_job_id` that preserves context between API calls, so the prior video does not need to be re-uploaded. Each request can use multiple reference images, including up to three per character for two characters, a motion reference clip of up to three seconds, and audio cues.

Extension is separate from native generation. You can append 3 to 10 seconds per turn, using the final 10 seconds as context, up to 30 seconds total. You cannot prepend footage or insert material mid-clip, and first-to-last-frame interpolation is still limited to outputs of 3 to 10 seconds. Those limits are worth knowing before someone promises a client a one-take 30-second commercial.

The audio is the part nobody prices

One thing to watch in the per-second numbers: they cover video tokens. Omni 1.2 Flash generates native audio with the video, dialogue, ambience and sound effects, and the model handles multiple speakers with separate conditioning.

Audio is where AI video usually falls apart in production. A clip with convincing motion and lip-sync that drifts by 200 milliseconds is unusable, and fixing it in post costs more than regenerating. If the audio is genuinely synced at generation time, the saving is not in the model bill, it is in the editor's afternoon.

Google also says lip-sync and dialogue clarity improved, and that different voices can be assigned accurately to different characters. Those are small claims in a press release and large claims in a workflow. Anyone evaluating this for advertising should test dialogue-heavy scenes first, because that is where the previous generation of models failed most visibly.

The migration date is the underrated detail

Omni 1.1 Flash stays available during a four-week transition. On December 3, API requests to 1.1 will route automatically to 1.2 without code changes, at the lower 1.2 rates.

For teams that built on 1.1, that is a free upgrade with one caveat: the model behind your product changes without your involvement. Any prompt library tuned against 1.1's behaviour, and any evaluation set that established quality baselines, now describes a model you are no longer calling. Re-run the evals in early December.

What it still cannot do

Google is candid about the gaps. Fast sports action, rapid script changes and complex physics remain under development. The Gemini API fully supports English and has not been evaluated for other languages. Voice editing, uploaded audio inputs and audio references are unsupported in the current API, with audio references gated behind allowlisting on Google Cloud.

The English-only limitation is the one that will bite international teams first. A German or Japanese campaign built on this model is running on a language the provider has not evaluated.

The real shift

For two years, AI video pricing was structured around the assumption that generation was expensive and therefore scarce. Every product decision followed from that: short clips, low resolution, aggressive prompt engineering, tolerance for bad output.

Omni 1.2 Flash prices generation as a commodity and puts the cost where it belongs, in the final render. Twelve cents to find out whether an idea works is a habit rather than a budget item.

Google also shipped the distribution in the same week: Flow for scene direction and multi-turn editing, Google Vids, the Gemini app for Ultra subscribers, and Google Ads campaign creation. When the model gets cheap and the surfaces are already installed, the constraint on AI video stops being access. It becomes whether anyone has a reason to watch.

Related articles