Google's Gemini Omni Flash 1.2 Cuts Video Costs 30% and Climbs Rankings
Google's newest video model climbs the Artificial Analysis boards with a 31% price cut, 15-second clips, and a #1 finish in complex edits.
- Gemini Omni Flash 1.2 is generally available on Gemini API and Google Cloud with native 15-second clips.
- Price drops 30% to $0.07 per second of 720p video, or $6.26 per minute at 1080p.
- Takes #2 on AA-Video-Editing, #6 on text-to-video, #14 on image-to-video.
- Renders a 10-second 1080p clip in about 76 seconds, 3x faster than Wan 3.0 or Seedance 2.5.
- Best at dialogue scenes, motion graphics, and text rendering; weak at human anatomy and live-action.
- API calls to
gemini-omni-1.1-flashauto-route to 1.2 starting December 3, 2026.
Google has made Gemini Omni Flash 1.2 generally available through the Gemini API and Gemini for business. On the Artificial Analysis leaderboards, the model rose from No. 12 to No. 6 for text-to-video, from No. 17 to No. 14 for image-to-video, and entered video editing at No. 2. The release adds longer clips and stronger editing capabilities while cutting 720p generation prices by 30% from Omni 1.1.
Longer clips, lower API costs
Gemini Omni Flash 1.2 extends the maximum clip length from 10 to 15 seconds and supports duration controls in one-second increments. Google's launch post also lists synchronized audio, output up to 4K, multi-turn conversational editing, and stronger consistency across as many as six reference images.
| Capability | Gemini Omni Flash 1.2 |
|---|---|
| Clip duration | 3 to 15 seconds, selectable in one-second increments |
| Audio | Synchronized sound generated with the video |
| Maximum resolution | Up to 4K |
| 720p price | $0.07 per generated second, or $1.05 for a 15-second clip |
| Reference images | Up to six, with improved subject and product consistency |
| Editing | Successive prompts can refine the same clip |
Consistency across reference images matters for applications that must preserve the same person, product, or visual identity through several shots. Multi-turn editing also lets an application revise a clip without rebuilding the entire prompt and workflow for each change.
One model ID, several surfaces
- Gemini API:
gemini-omni-1.2-flash - Developer tools: Google AI Studio and Vertex AI
- Google products: Gemini, Flow, YouTube, and Google Ads
- Partner tools: Figma's Weave and Magnific, with Adobe planning Firefly support
Starting December 3, 2026, requests to gemini-omni-1.1-flash will route automatically to version 1.2 and use its billing rates, according to the Gemini API changelog. Existing applications can continue using the old model ID, but their output behavior and costs will change. Teams with regression tests or fixed quality targets should validate 1.2 before the cutoff.
Ranks rise across three tasks
Text-to-video creates a clip from a written prompt, image-to-video animates a supplied starting image, and video editing modifies existing footage. Omni 1.2 improved its position on all three Artificial Analysis boards.
| Task | Omni 1.1 | Omni 1.2 |
|---|---|---|
| Text-to-video | No. 12 | No. 6 |
| Image-to-video | No. 17 | No. 14 |
| Video editing | No prior overall rank cited | No. 2, Elo 1,164 |
An Elo score expresses relative performance in head-to-head comparisons. Omni 1.2's editing score of 1,164 trails Wan 3.0 at 1,192. The score is useful for comparing models evaluated on the same board; it does not represent a percentage quality grade. Rankings can change as new models and evaluations are added.
A 10-second 1080p clip rendered in about 76 seconds, faster than every model ranked above it on the text-to-video board. MiniMax H3 at 768p took about 97 seconds for a five-second clip, while listed runs for Wan 3.0 and Dreamina Seedance 2.5 approached four minutes. The differing resolutions and clip lengths make these directional throughput comparisons rather than controlled latency tests.
Google's 720p list price of $0.07 per second equals $4.20 per generated minute. Artificial Analysis reports $6.26 per minute for its leaderboard configuration, which uses different generation settings. MiniMax H3 at 768p costs $4.80 per minute and ranks higher; Omni 1.2's leaderboard cost remains below most models on the board. Developers should estimate production costs from the API resolution, duration, and retry rate used by their application.
Dialogue and design set the pace
- Dialogue and lip sync: No. 1, up from No. 7
- Multi-scene narrative: No. 2 for stories spanning several shots
- UI, UX, and motion design: No. 1, with leading results in retail, ecommerce, animation, and marketing
- Audio synchronization: No. 2, behind Omni Flash 1.1
Starting from an image, Omni 1.2 improved in 19 of 20 measured capabilities. Text rendering climbed from No. 13 to No. 3, indicating better preservation of readable lettering during animation. Audio synchronization reached No. 5.
In video-editing tests, the model moved from No. 2 to No. 1 for complex edits and from No. 3 to No. 1 for visual-effects editing. It finished ahead of Wan 3.0 in both categories.
Motion exposes the gaps
Realistic footage of moving people remains the weakest area. Omni 1.2 ranks No. 8 in live-action film, No. 9 in consumer content, and No. 12 in human anatomy for image-to-video. Its stronger human results concentrate in dialogue and lip sync, while anatomy and full-body movement rank lower.
Sound and speech editing is its lowest-ranked editing action at No. 8, down from No. 7 for Omni 1.1. Physics and world simulation also slipped from No. 4 to No. 6, limiting its fit for scenes that depend on precise collisions, weight, or complex physical motion.
Where it fits the stack
| Workload | Assessment |
|---|---|
| Dialogue scenes and presenters | Strong fit based on the No. 1 dialogue and lip-sync result |
| Multi-shot narratives | Strong fit, with a No. 2 ranking for multi-scene storytelling |
| Motion graphics and animated interfaces | Strong fit based on its leading UI, UX, and motion-design scores |
| Product and marketing clips | Strong fit when reference consistency and readable text matter |
| Cinematic live action and athletic movement | Evaluate Wan 3.0, Seedance 2.5, and MiniMax H3 alongside it |
| Physics-heavy scenes and speech editing | Run targeted tests because these categories declined from version 1.1 |
A 15-second maximum matches a common social-ad slot without stitching multiple clips. At the 720p list price, each full-length generation costs $1.05 before retries, and the reported 76-second render time provides a useful baseline for capacity planning. For teams producing ad variants, those duration, cost, and latency figures directly determine iteration throughput.
Comments
