Gemini Omni Flash 1.2 Leak: One Quota Row, One Video, and Nothing in Google's Docs

The entire public record of Gemini Omni Flash 1.2 on 5 November 2026 is a screenshot of a Google Cloud quota page and a ten-second video. That is it. Google's developer documentation lists Gemini Omni 1.1 Flash and nothing else in the Omni line — not on the model list, not on the pricing page, not in the changelog, not in the Vertex AI catalogue — so what follows is a what-we-know-so-far piece about a version that has been sighted and not shipped. Every dollar figure in it is either a rate Google publishes today for 1.1 Flash or somebody's unverified claim about 1.2, and the two are kept apart throughout.

The sightings arrived close together. On 5 November the X account @chetaslua posted a clip it attributes to the newer model and called "minor update over flash 1.1." Around the same time @mrfanduu posted a Google Cloud screenshot naming a model ID in the version-1.2 shape and, separately, the claim that it runs roughly 30% cheaper than 1.1 Flash. TestingCatalog folded the cloud sighting into its daily brief the same day with the line "Gemini Omni 1.2 Flash was spotted on Google Cloud (via mrfanduu)." None of the three is Google, and none of them includes a model card, a price table, a spec sheet or a date.

The evidence, item by item

Three artifacts exist, and they are not of equal weight. Sorting them is most of the work of reading this story correctly.

• The cloud screenshot. A quota or service listing in the Google Cloud console carrying an Omni model ID at version 1.2, posted by @mrfanduu and relayed by TestingCatalog. This is the strongest artifact of the three, and it is still a billing surface rather than a product surface — more on what that does and does not prove below.

• The price claim. The same account's assertion that the model is about 30% cheaper than 1.1 Flash. There is no published rate behind it, no currency, no tier and no unit — a percentage and nothing else. Treat it as a rumour with a number attached, which is the most seductive and least checkable kind.

• The video. @chetaslua's clip, described as generated by Omni Flash 1.2 and characterised as a minor update to the previous version. The poster says the clip has the model explaining pricing using Dragon Ball characters. It carries no verifiable provenance — no model ID in a request header, no API response, nothing but the poster's word — and a single clip cannot establish what a model does in general.

One structural point about all three: they are sightings of different things. The cloud row is evidence that plumbing exists somewhere in Google's stack. The price claim is evidence that somebody believes something. The video is evidence that a video exists. Only the first is a fact about Google's infrastructure, and even that is a fact about a quota entry rather than about a model.

What Google's own pages say, checked today

The falsification is straightforward and takes about four pages. On 5 November 2026:

• The Gemini API model list names exactly one Omni model, Gemini Omni Flash, with the ID gemini-omni-1.1-flash. No 1.2 entry, no preview alias, no "coming soon" row.

• The Gemini API pricing page prices one Omni model: gemini-omni-1.1-flash, billed as a video generation and editing model, with no free tier. No second Omni line item appears anywhere on it.

• The Vertex AI model catalogue groups "Gemini Omni 1.1 Flash" and "Gemini Omni Flash" under preview models. Neither is a 1.2, and the second is the older preview ID that 1.1 replaced.

• The changelog is current — it carries entries dated 5 November 2026 for Office document inputs and 4 November for Gemini 3.8 TTS voice remixing — and its most recent Omni item is still 27 August 2026, the day gemini-omni-1.1-flash went generally available. A release that shipped would move that line. Nothing has moved it.

Google's deprecation table completes the picture: the older gemini-omni-flash-preview ID is retired with 1.1 Flash named as its replacement, and gemini-omni-1.1-flash itself carries no shutdown date. Google is, by its own documentation, still selling 1.1 as the current Omni model.

Why a cloud quota row is a real signal, and a weak one

Quota entries appearing ahead of an announcement is a known pattern on Google Cloud, and it is not nothing. Serving infrastructure and billing meters get provisioned before a model becomes callable; the console can show a quota line for something the API will not answer for weeks, and users have reported exactly this for earlier Gemini Flash generations before their launches. If you are tracking a model's approach, a cloud row is one of the earliest surfaces it can touch.

What it cannot tell you is nearly everything a buyer needs. A quota row has no context window, no clip length, no resolution ladder, no price per second, no availability tier and no launch date. It is also reversible: provisioned capacity that never becomes a launch is ordinary, and a console line is not a commitment. So the correct reading is "Google has built or is building something under this ID," not "this model is coming out." Note also the shape of the claim — the ID in circulation is gemini-omni-1.2-flash, which is consistent with the naming Google already uses, which is exactly what someone fabricating a screenshot would also produce. A consistent ID is evidence of consistency, not of authenticity.

The 30% number, priced out

The one claim with a figure in it is worth taking seriously enough to do the arithmetic — not because it is verified, but because the shape of the answer tells you how much the claim would be worth if it were true.

Gemini Omni 1.1 Flash is billed on tokens, not seconds. Google's pricing page lists video output at $17.50 per million tokens, and documents 5,792 tokens per second of 720p video, which works out to roughly $0.10 per second — the figure everyone quotes. Input is $1.50 per million tokens whether the input is text, image, video or audio, and text output is $9.00 per million; against a video output stream those are small numbers, and the bill is set by output tokens.

Apply a 30% cut to that video output rate and you get $12.25 per million tokens, or about $0.071 per second of 720p:

• A ten-second clip — $1.01 today, about $0.71 at a 30% cut.

• A forty-second scene built by extension — $4.05 today, about $2.84 at a 30% cut.

• A hundred ten-second clips a month — $101 today, about $71.

That is a meaningful line for anyone generating video at volume, and a rounding error for anyone generating a handful of clips a week. It is also arithmetic on top of an unverified percentage: if 1.2 ships, it may not be 30% cheaper, the cut may apply to one tier and not another, and Google has not published a rate for 1080p or 4K on 1.1 at all, so there is no published baseline for a discount to be measured against at the resolutions most production work uses. Hold the number loosely.

The same discipline applies to whatever you actually pay: vendor list price is the floor, and platforms bill on top of it. OrcaRouter passes provider list price through with no markup, so a Google price change is visible here on the day it lands rather than after a margin layer absorbs it — but that only matters for models we route, and we do not route this one. More on that in a moment.

What "minor update" has to mean

@chetaslua's phrase is the most useful part of the leak because it sets the expectation to check against. A point release on a video model is not a small thing if it touches any of the constraints that decide whether the model is usable. Here is what 1.2 would be measured against, drawn from Google's own documentation and the partner integration notes for 1.1 Flash:

• Clip length. Generates 3–10 seconds per request. The ten-second ceiling is the most-cited limitation of the whole family, and it is the first thing a "minor update" would have to move to matter.

• Extension. Continues a scene in roughly ten-second steps to about 40 seconds total, reading a longer window of prior footage than the original preview did.

• Controls. First and last frame keyframes, up to 14 reference images, and up to three reference video clips of ten seconds each in a single call.

• Resolution and shape. 360p draft tier at roughly one third the cost and meaningfully faster (both Google's figures, unreproduced), then 720p, 1080p or 4K, in 16:9 and 9:16.

• Interface. Multimodal input — text, image, audio, video together — conversational multi-turn editing through the Interactions API, and native synchronised audio on every clip.

• Provenance. Every clip carries a SynthID watermark and an AI label.

The gaps a 1.2 could plausibly close are visible in that list. The per-second rate for 1080p and 4K has never been published, and a higher-end Omni Pro tier is reported to be in the works with no release date, which leaves the Flash line as the only Omni that ships. A .2 bump ten weeks after 1.1 — the preview landed 30 June, 1.1 arrived 27 August — fits a cadence of small, frequent updates rather than a generational jump, which is consistent with "minor" and inconsistent with everything people actually want from the line.

There is also no independent benchmark for any Omni model. We could find no Omni entry on Artificial Analysis's model index, and the arena leaderboards we checked list no Omni row — and every quality claim about 1.1 in circulation traces back to Google. That matters for 1.2 in a specific way: if a new version ships, the first honest signal about it will not be a vendor number, and it will not be a single generated clip either. It will be a third-party comparison under a fixed prompt set.

The Dragon Ball detail is the part worth watching

Buried in a post about version numbers is the most interesting claim in the leak: that the clip shows Dragon Ball characters explaining pricing. If that is what the model produced from a prompt, three things follow, and only one of them is about pricing.

The first is instruction-following at length. A ten-second clip that carries a structured explanation — characters, a topic, and numbers — is a harder task than a pretty landscape, and it is the kind of task that separates a demo model from a production one.

The second is on-screen text. Rendering legible numerals inside generated video has been a weak point for every model in this class, and a clip explaining a price list is a direct test of it. The poster's framing implies the model put the pricing on screen, though the post itself does not say whether the numbers it produced were correct, which is the only part a buyer would care about.

The third is intellectual property, and it is the one Google has to answer for rather than the poster. Recognisable characters from a licensed anime franchise appearing in a generated video is exactly the boundary every video model has been negotiating, and a model that draws them on request is a model whose guardrails will be scrutinised the week it launches. None of that is established here: we do not know the prompt, whether reference images were supplied, whether the clip was generated or edited from source footage, or whether the characters are faithful or merely suggestive. A single unlabelled clip proves that a video exists. It does not prove what the model did to make it.

What to do between now and the launch

Nothing in your stack needs to change today, because nothing has shipped. Gemini Omni 1.1 Flash remains the model Google documents, prices and supports, and it remains the one to build against.

Four signals are worth watching, and they are ranked by how much they would settle. A model ID appearing in the Gemini API model list or a pricing line on the pricing page would be a launch. A changelog entry dated after 27 August would be a launch. A third-party evaluation under a fixed prompt set would be the first evidence about quality. And a shutdown date on gemini-omni-1.1-flash in the deprecation table would tell you Google intends to move everyone — that is the one that forces a decision rather than inviting one.

On routing, the honest state of things: Google has not opened the Omni video API to third-party routing, OrcaRouter carries no Omni model, and there is no OrcaRouter model page for either 1.1 or 1.2 — we checked, and both return 404. We are not going to imply otherwise. What we do carry is the rest of Google's line: Gemini 3.8 Flash and Gemini 4 Argon are both callable here, alongside more than 200 other models on one API key, one bill. And when a model we do carry gets a price cut, the new rate is live on our side the same day — provider list price passed through, no margin layer in between. If Omni ever becomes routable, the 30% question becomes a number you can check rather than a percentage in a post.

What would turn this from a leak into a launch story is small and specific: one model ID on a documentation page, or one price row on a pricing page. Until one of those appears, the record is a quota entry and a clip, and the useful thing to hold onto is the gap between them — how much of what a buyer needs is missing from even a convincing sighting.