News · 2026-11-07

Google expands Gemini Omni Flash video generation to 15 seconds

Google has released Gemini Omni 1.2 Flash with direct video generation and editing for clips up to 15 seconds. Its documented task limits differ: interpolation and extension generate segments of up to 10 seconds, while video understanding accepts longer inputs.

Key facts

• Google Cloud lists direct generation and editing at 3–15 seconds.
• First/last-frame interpolation and extension are limited to 10-second generated segments.
• Google lists 720p generation at $0.07 per second through its API.
• The Google announcement and Cloud documentation describe the release.

The practical difference is between creating a clip, modifying an existing one and inspecting a longer video. Google Cloud’s capability table specifies limits for those separate tasks. Its extension guide describes appending a generated segment to part of an existing clip, so an assembled output can be longer than the new segment alone.

Google describes a system conditioned on combinations of instructions and references, with video and synchronized audio as output. A useful analogy is a creative brief with reference images: the references constrain what the system should make, but they do not guarantee that every object or detail remains unchanged. Its model card describes supported modalities and safety measures, while the Cloud launch says audio references require allowlisting.

The API also supports stored interactions for multi-turn edits, described in the editing guide. That can let a user continue an editing session without rebuilding the whole request. Google says retained interactions are available for up to seven days; users handling sensitive material should understand the product’s retention settings.

Google calls the model “Generally available” in Cloud. That does not mean every feature or consumer interface has the same entitlement. At the stated 720p API rate, 15 seconds costs about $1.05 before any other relevant charges; consumer credits are a separate meter.

The discussion includes users who value longer shots and others who find continuity or direction-following unreliable. The company provides examples but no 1.2-specific public video-quality benchmark in the reviewed card. Its examples are therefore evidence of possible outputs, not a measured guarantee of fidelity, speed or superiority across tasks.

The feature list also contains a distinction between what a model can theoretically accept and what an individual account can use. Google Cloud's description of audio references includes an allowlist. The model card also says speech alteration in edits is restricted while the company studies safe deployment. A developer building a workflow should therefore check the exact product and account configuration rather than promise an end user every modality listed in a general model description.

The interaction design changes the economics of revisions. A stored conversation can refer back to earlier work, which may reduce the need to reconstruct a brief. It can also create obligations around retained inputs and outputs. The Cloud editing guide describes the interaction API and its stateful and stateless options. A stored session is useful precisely because the service retains context; that retention needs to be considered for sensitive footage and customer data.

Quality is harder to summarize than duration. A 15-second clip may have convincing movement but inconsistent lettering, an altered product or a character that changes between frames. A shorter generated segment anchored to two reference frames can be more useful to a creator than a longer unanchored output. Google shows examples and customers describe use cases, but the reviewed model card does not supply a public, 1.2-specific blind comparison with sample sizes and error rates.

Price comparisons have similar traps. At the listed API rate, a single 15-second 720p generation is about $1.05, but a finished shot may require multiple attempts. Consumer Flow credits are not dollars per API second, and reports in the discussion use different settings and accounts. The meaningful business measure is the cost and time per usable result under a controlled workflow.

Primary source, verified: read the paper →

Ground Truth. (2026, November 7). Google expands Gemini Omni Flash video generation to 15 seconds. Ground Truth. https://groundtruth.day/news/google-gemini-omni-1-2-flash-video.html

@misc{groundtruth:google-gemini-omni-1-2-flash-video,
  title  = {Google expands Gemini Omni Flash video generation to 15 seconds},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {nov},
  url    = {https://groundtruth.day/news/google-gemini-omni-1-2-flash-video.html}
}

Topics: google · video · product