Gemini Omni 1.2 Flash's spec sheet is short, but several details are easy to trip over: 1080p and 4K are upscaled, not native; aspect ratios are only two; first/last-frame interpolation caps at 10s; extension is append-only. This article walks through the model page and developer guide (read directly Nov 14, 2026) item by item.
Generation and editing 3–15s, whole-second steps, default 10s (1.1 capped at 10s). Resolutions: 360p, 720p, 1080p, 4K — 1080p and 4K are upscaled outputs, not native high-resolution generation. Aspect ratios only 16:9 (default) and 9:16 — no 1:1, 4:3 or 21:9. Output is MP4 at 24 FPS. First/last-frame interpolation (two images as first and last frame) only supports 3–10s; over 10s returns an invalid-argument error.
Up to 6 reference images (all durations). Reference video up to 3s — up to 3 for 3–10s videos, only 1 for 11–15s; reference video only for new generation, not editing or extension; audio in reference video is ignored. Compared with Seedance 2.5 (up to 50 reference assets per generation: 30 images + 10 videos + 10 audio), Omni's reference capacity is much smaller.
Conversational editing goes through the Interactions API, referencing the previous round with previous_interaction_id; you can change background, lighting, weather, aspect ratio, swap products and more; the model remembers the earlier video context and preserves unmentioned elements; each round produces a new video and consumes a generation + output tokens. Extension adds 3–10s per turn (default 10s), up to 30s total across rounds; append-only at the end (no prepend or mid-insert); extension retouches the last frames for a smooth transition. Note: 1.1's cumulative extension cap was 40s, higher than 1.2's 30s — an easy-to-miss downgrade at migration. JoinOasis tested it: a 10s clip + extension = 20s video, seam consistent, but an unwanted dissolve appeared ~3s into the extension.
Interactions API only (POST https://generativelanguage.googleapis.com/v1beta/interactions), no generateContent; SDK client.interactions.create; large videos (>4MB, ~above 720p) should use delivery="uri". Token caps: input 131,072, output 65,536 (same for 1.1 and 1.2). Rate limits: free tier not supported; Tier 1: 8 RPM / 40 RPD; Tier 2: 16 RPM / 200 RPD; Tier 3: 40 RPM / 2,000 RPD; the changelog says 1.2's paid limits are 2x 1.1's. Each edit/extension round is an independent interactions.create call that consumes one request quota. No Batch API pricing; the Gemini API doesn't support Provisioned Throughput (Google Cloud does).
The developer guide's Limitations section (read directly Nov 14, 2026): EEA, Switzerland and UK users currently cannot edit or extend uploaded video (editing/extending model-generated video still works there); the same regions also don't support uploading/editing images containing minors; images with certain identifiable people are unsupported for upload/edit. Voice editing not supported; the API doesn't support audio input or audio references (Cloud blog: audio references need allowlisting). Cannot extend an uploaded video of a speaking person to add dialogue (keeping the character silent or without added dialogue is OK); multi-round extension of model-generated video supports dialogue generation. Cannot reference/reason across multiple videos; YouTube videos can't be a media source. No system instructions, temperature, top_p, stop sequences or negative prompts (you can write "don't do X" in the prompt). English is the only fully supported language; other languages are "not yet evaluated, results may vary". Officially admitted weak spots (blog): fast sports action, fast script changes, complex physics.
All generated videos carry an invisible SynthID watermark (programmable detection); prompts and generated videos pass content-safety filtering (varies by region). Segmind's test adds: the copyright filter rejects outright when source material contains real brand logos even if the prompt doesn't mention the brand. QuantSlant's test: prompts involving children were rejected and not charged.
Reference: Google — Generate and edit videos with Gemini Omni Flash (developer guide) (rewritten and organized by ModelBrief; facts per the official source)