News · · 8 min read

Gemini Omni 1.2 Flash Goes GA: 1.1 Calls Reroute on December 3, and the Endpoint Is Global-Only

In this post
  1. The December 3 cutover for 1.1
  2. 1.1 and 1.2 on the same rows
  3. Length limits by mode
  4. Why there are no performance numbers yet
  5. Where you can use it, and on what terms
  6. Hacker News on whether 15 seconds matters
  7. What to act on now
  8. Limitations
  9. Sources

Google made its video generation and editing model Gemini Omni 1.2 Flash (gemini-omni-1.2-flash) generally available on the Gemini API on November 5, 2026. A single clip can now run 15 seconds instead of 10, and Google says the price is 30% lower than 1.1. For developers already shipping on 1.1, though, the bigger change is a deadline: starting December 3, every request to gemini-omni-1.1-flash is routed to 1.2 with no code change. And on Google Cloud, the model is only offered on the global endpoint, which carries no guarantee about where a request is processed.

This post covers what changes between 1.1 and 1.2, the length limits that differ by mode, why there are no performance numbers yet, the access paths and their conditions, and the arguments for and against on Hacker News. We did not generate any videos ourselves. Every comparison and figure here is compiled by us from Google's official documentation.

The December 3 cutover for 1.1

If you already call 1.1, you have a fixed window. The Gemini API docs keep gemini-omni-1.1-flash available for a 4-week transition only. From December 3, 2026, all 1.1 requests automatically go to gemini-omni-1.2-flash and are billed at 1.2 rates.

The cutover will not cost you more. It will, however, change your outputs even if you pinned the model name. A service that runs the same prompts over and over, such as product or ad videos, may see different color, motion, or delivery after December 3. The docs offer no way to stay on 1.1. The practical move is to run your current prompts against 1.2 now and compare the results before the switch happens for you.

1.1 and 1.2 on the same rows

We pulled matching items for both models from Google's blog and the Gemini API pricing page:

ItemOmni 1.1 FlashOmni 1.2 Flash
Clip length per generationUp to 10 s3–15 s, in 1 s steps
Video output price (per 1M tokens)$17.50$12.00
Effective price per second, 720pAbout $0.10About $0.07
API statusRoutes to 1.2 from Dec 3Generally available

Both models count video the same way: one second of 720p is 5,792 output tokens. On that basis the pricing page puts a 15-second 720p clip from 1.2 at about $1.04. The same page lists per-second figures for other resolutions: about $0.023 for 360p, $0.104 for upsampled 1080p, and $0.209 for upsampled 4K. Input costs $1.50 per million tokens for both models, and neither is available on the free tier.

Beyond length and price, most of what changed is a quality claim. The Google Cloud blog says 1.2 follows image and video references more closely, so products and characters stay consistent from shot to shot, and that dialogue sounds more natural with clearer lip-sync. Google also leans on a draft workflow: generate at 360p or 720p, then upscale only the keepers to 1080p or 4K. As the performance section below explains, none of these quality claims has a number behind it yet.

Length limits by mode

"15 seconds" is not the cap for every mode. Laying out the constraints from the Gemini API docs by mode gives Figure 1.

Figure 1. Video length ranges by mode for Gemini Omni 1.2 Flash, compiled by us from the Gemini API Omni docs and Google's blog (November 5, 2026). For 1.1, only the published 10-second generation cap is shown.

First and last frame mode is the first wall you hit. If you supply both a starting and an ending image, the clip can only be 3 to 10 seconds long. Asking for more than 10 returns an invalid argument error. If your workflow pins how a character looks at the start and end of a shot, the jump to 15 seconds does nothing for you.

Reference videos work only when generating a new clip, and their limit depends on length. A 3–10 second clip accepts up to three reference videos of 3 seconds or less, while an 11–15 second clip accepts only one. Reference images are capped at six regardless of length. Extension only appends to the end of a clip, 3 to 10 seconds per turn, up to 30 seconds in total.

Why there are no performance numbers yet

Right now nobody can put a number on 1.2. Google's blog publishes no benchmark scores and shows the difference from 1.1 only through example clips. It openly says work is still ongoing on fast sports action, rapid script changes, and complex physics.

There is no third-party measurement either. When we checked the Artificial Analysis video leaderboards (text-to-video, image-to-video, and video editing) on November 7, the only Omni entry was 1.1. At that time 1.1 had a text-to-video Elo of 1,127 across 7,807 appearances. How far 1.2 moves from that score will likely be the first independent check on Google's quality claims.

Where you can use it, and on what terms

The conditions differ by access path. We kept only what Google's own docs and blog state:

PathConditionsWhat to check
Gemini API (AI Studio)Paid tier only; Interactions API onlySouth Korea is on the supported list
Google Cloud (Gemini Enterprise Agent Platform)global endpoint only; Provisioned Throughput supportedNo regional processing guarantee
Gemini app15 s video creation rolling out to Google AI Ultra subscribersNo per-country schedule published
Google Vids15 s clips at no extra cost with a Google account or WorkspaceNo per-country schedule published

The API surface is the first thing to notice. According to the Google Cloud model page, 1.2 does not support the GenerateContent API or the Chat Completions API; it works only through the Interactions API. System instructions, structured output, function calling, and batch inference are also unsupported. You cannot take code that calls a Gemini text model and swap in the model name.

For teams with data-location rules, the endpoint matters more. The Google Cloud model page lists a single supported region: global. Google Cloud's data residency page says global endpoints "route and process data anywhere globally, without restricting it to a specific geographic region," and do not provide data residency guarantees. Stored data stays in the location you chose, but you cannot choose where the model processes a request. If your service sends customer photos or unreleased product images, it appears you will need to confirm first whether your security policy requires in-country processing.

Audio references differ by path too. The Google Cloud blog says audio references are gated behind allowlisting, while the Gemini API docs say uploading audio inputs or audio references is unsupported in the current version of the API. Matching dialogue to a recorded voice is not something you can do out of the box on either path today.

Hacker News on whether 15 seconds matters

The November 6 Hacker News thread on Google's announcement drew many comments about Gemini plans, access paths, and naming in general rather than the Omni model itself. On the model, opinions split over whether 15 seconds actually matters.

The positive case came from a filmmaking angle. User msabalau wrote:

"Increasing the available range of shot lengths from 10 seconds to 15 seconds gives film makers a wider pallet of options, if the longer shots have continuity, which is the other claim here."

In the same comment, they noted that the average shot in a feature film runs about three seconds, so for one-and-done use the gap between 10 and 15 seconds may be small.

The skeptical case pointed at the first/last frame limit covered above. User fuzzythrowaway2, who says they have been making AI video since early 2024, wrote:

"Not being able to use first and last frames for the longer videos doesn't render it useless, but it gets pretty close"

They added that without first and last frames, unusual character designs "regress to tv commercial blandness," and that they would rather wait for a clip that follows directions than get a fast one that does not.

On price, one user observed the two paths moving in opposite directions. User msabalau reported that in Flow, Google's video tool, a six-second video shows 10 credits on 1.1 and 18 credits on 1.2. In other words, the API got cheaper while Flow credits went up. That is one user's on-screen observation, not a figure Google has published.

What to act on now

Two things are worth doing now. If you run 1.1 in production, test the same prompts on 1.2 before December 3. If you plan to use it through Google Cloud, confirm first that the global endpoint passes your data policy.

Fifteen seconds in vertical 9:16 looks like a good fit for short-form ads and product clips. Dialogue in other languages is a risk too: the Gemini API docs say English is fully supported but other languages "have not been evaluated, so they may work but results can vary." Keyframed shots are still capped at 10 seconds, so in many cases a plan built on "10 seconds or less, then extend" may work better than "one 15-second shot." There is no free tier, so budget for the paid tier even for testing.

Limitations

ItemDetail
Hands-on generationWe did not generate any videos for this post. All quality statements are Google's claims or user opinions
Performance numbersGoogle published no benchmarks, and as of November 7, 1.2 is not on the Artificial Analysis leaderboards
Leaderboard valuesThe 1.1 Elo is a snapshot; leaderboards keep moving
Dialogue languagesThe Gemini API docs say only English is fully supported and other languages have not been evaluated. We could not check Korean dialogue quality
Gemini app and VidsWe found no official per-country rollout dates
Flow creditsFrom a single Hacker News user; not confirmed against an official price list

We covered a similar forced move on the text side in what changes when you migrate to Gemini 4 Argon.

Sources

1 / 1