Runway Integrates Gemini Omni 1.2 Flash, Pushing AI Video Clips to 15 Seconds

Creators can now use Google DeepMind’s Gemini Omni 1.2 Flash inside the Runway editor . The model generates clips up to 15 seconds with synchronized audio, follows visual references, and supports iterative scene edits. Runway can then upscale selected footage to 4K.

For teams testing AI video pipelines, the integration removes several manual handoffs. Creative users can generate, revise, compare, upscale, and review footage on one timeline. Developers building server-side products still need to confirm whether Runway exposes the model through its API or use Google’s API directly.

The revision loop gets shorter

Gemini Omni 1.2 Flash raises the maximum generation length from the 10 seconds available in Omni 1.1 Flash to 15 seconds, a 50% increase. Google also cut the listed 720p price by 30%, from about $0.10 to $0.07 per generated second.

Multi-turn editing lets the model carry a clip and its instruction history across successive revisions. A creator can request, Keep the shot and change the lighting to golden hour, then continue refining the result without submitting the full brief again or regenerating the original scene.

Inside Runway, that process becomes a compact production loop:

1. Generate or import a base clip.
2. Revise lighting, objects, framing, dialogue, or camera treatment.
3. Compare variants on the timeline.
4. Upscale the selected take after the edit is locked.

Inputs and limits, mode by mode

Omni 1.2 Flash supports generation and editing with synchronized audio. Visual references can establish a character, product, setting, or style, while text instructions specify dialogue and sound. Uploaded voice or music references are unavailable, so audio matching remains outside the model’s reference workflow.

Mixed-input jobs support up to five images and three videos, with motion references limited to clips of roughly three seconds. Limits can vary by mode and interface, so automated workflows should validate file count, duration, format, and size before upload.

The $1.05 baseline

Google lists 720p generation at $0.07 per second, which puts a full 15-second clip at $1.05 before extensions, edits, or higher-resolution processing. Runway converts usage into workspace credits, so the effective price depends on the plan, generation mode, resolution, and any separate upscaling step.

Low-resolution drafts offer the clearest cost control. Teams can iterate at 360p or 720p, approve a take, and spend additional credits on 4K upscaling only for selected footage.

Existing footage gets the clearest gains

The model suits projects that begin with usable footage and require controlled variations. Advertising teams can change products, lighting, or framing for different placements, while narrative teams can refine dialogue and atmosphere without rebuilding every shot.

Native synchronized audio removes a separate sound-generation pass from early drafts. The absence of uploaded audio references limits voice matching, music matching, and sound-led edits, while the 15-second ceiling means longer sequences still require extensions or timeline assembly.

Every generated clip includes Google’s SynthID provenance watermark. Teams with detection, disclosure, archival, or compliance requirements should account for that marker in downstream processing.

Checks before wiring it into production

Runway’s announcement establishes interactive access inside the editor. Server-side availability remains a separate implementation question, especially for products that need queued jobs, automated asset handling, or generation at scale.

• Confirm the exact model identifier and version-pinning policy.
• Check whether Runway exposes the model through its API.
• Record rate limits, concurrency limits, timeouts, and retry behavior.
• Validate supported codecs, file sizes, durations, and reference counts.
• Measure credit use for edits, extensions, failed jobs, and upscaling.
• Review asset retention, training, privacy, and regional processing terms.
• Test continuity across repeated edits, especially for faces, products, and dialogue.
• Preserve SynthID and other provenance metadata through export and transcoding.

One model, several workflow layers

Artlist, FLORA, InVideo, Krea, Picsart, Photoroom, and QuickFrame AI also offer the model through their creative products. That distribution gives teams several interfaces for the same underlying generation system, making workflow design, governance, collaboration, and pricing meaningful points of comparison.

Teams already using Runway’s Gen-4 stack can evaluate both models on the same timeline with matched prompts and references. That setup supports direct comparisons of motion, continuity, dialogue, editability, latency, and cost without building a separate Google API pipeline.