Google launched Gemini Omni at I/O 2026 last Monday, and we spent the week running the same prompts through it, OpenAI’s Sora, Runway Gen-4, and ByteDance’s Seedance 2.0. The standings were tighter than expected, and the winner was not the prettiest.
The setup
Monday night, Sundar Pichai and Demis Hassabis opened Google I/O 2026 with Gemini Omni. Hassabis pitched it as a world model, not a text-to-video tool. We wanted to find out whether that distinction holds up in practice, or whether it’s a press-release flourish.
So we picked four contenders. Gemini Omni Flash, launched five days ago. OpenAI’s Sora, the incumbent in consumer AI video. Runway Gen-4, the pro-leaning tool that working filmmakers actually pay for. And ByteDance’s Seedance 2.0, the underdog that quietly out-benched everyone earlier this spring.
Same prompts, same expectations, four very different results.
The test that mattered
We started with one prompt across all four platforms: a street in Marseille in the rain, seen from the window of a café, late afternoon light. Standard scene, achievable by everyone.
Then we made it harder. Add a black cat crossing the street. Change the light to a sunset. This is the kind of iterative request creators actually make. Not a single prompt fired and forgotten, but a scene that evolves over three or four turns.
This is where the four split.
Gemini Omni: the conversational advantage
The Marseille clip itself was solid. Rain density was right, reflections on the asphalt moved correctly, pedestrians had believable trajectories. Nothing the others can’t do at this stage.
The cat is where Omni separated itself. We did not have to regenerate. We typed the instruction in chat, Omni took the existing clip, added the cat, kept the lighting, kept the rain, kept the pedestrians moving on the same paths. The sunset change worked the same way. New sky, same scene, no break in continuity.
The moment it clicked.
Three iterative edits, zero regenerations from scratch. Sora and Runway can’t do this today. Seedance can’t either. This is what Google means when it calls Omni a world model: the model holds the scene as a coherent space and edits it, instead of producing a brand-new clip every turn.
For a creator working on a single piece of content, this turns a two-hour session into twenty minutes. The clip cap is still 10 seconds at launch, which is a real limit. The conversational loop more than makes up for it on day-to-day work.
Sora: still the photorealism champion
Sora gave us the most cinematic Marseille frame of the four. Better lens character, better shadow gradation, the kind of image that holds up at 4K. If we had stopped at the first prompt, Sora would have won the round.
The iterative test broke Sora’s flow. Adding the cat meant a fresh generation. The rain density shifted, the pedestrians took new paths, the framing drifted slightly. Sora is producing a new clip each turn, not editing the one we already have. For a single hero shot that is fine. For a director chasing a specific composition across edits, it is friction.
Sora still wins on duration. Its 60-second clip cap is six times Omni’s. For anything beyond a social-format snippet, that gap matters.
Runway Gen-4: the pro tool with a pro learning curve
Runway gave us the most control. Motion brushes, camera-path overrides, frame-by-frame timing. Working filmmakers pay for Runway for a reason. The Marseille shot took longer to set up, but we could direct it in ways neither Omni nor Sora allow.
The cost is the learning curve. We spent well over an hour inside Runway to produce what Omni gave us in twenty minutes. For a VFX artist, that trade is obvious. For anyone else, it is a wall.
Runway also has no consumer distribution. There is no YouTube Shorts integration, no native social pipeline. You make the clip, you export, you upload. Omni’s distribution is structurally faster, and that advantage compounds.
Seedance 2.0: the underdog you can’t reach
Seedance produced our best raw image quality, by a thin margin. Skin texture, water surfaces, fabric folds. On a still-frame test, Seedance won. Independent benchmarks earlier this spring had it ahead of Veo 3.1 and competitive with Sora on raw fidelity.
The problem is access. Seedance lives inside ByteDance’s Doubao app and a handful of Western reseller wrappers. Getting a clean account, paying, and running a serious test is genuinely complicated outside China. For a Chinese creator this story reads differently. For everyone else, Seedance is currently a benchmark you read about, not a tool you ship with.
The verdict
If we had to recommend one model right now, based on a week of real use, it would be Gemini Omni Flash.
Not because it produced the best individual frames. Sora and Seedance both edged it on raw quality, and Runway gave us more direct control. Three of the four beat Omni on something.
Omni won on the thing that adds up across a real workday. Iterative direction actually works. And Google did something nobody else has: it shipped a frontier video model directly inside YouTube Shorts, free, on day one. The combination of conversational editing plus instant distribution is harder to dismiss than the spec sheet suggests.
What to use when
After a week of moving prompts between the four platforms, the cleanest read is to match the tool to the job.
- For Shorts creators and social-first work: Gemini Omni Flash inside the YouTube Shorts app. Free, fast, integrated end to end.
- For one-shot cinematic hero clips: Sora. Best individual frame quality and the longest clip length on the market.
- For directed work where every frame matters: Runway Gen-4. The professional pipeline, if you have the patience to learn it.
- For benchmarking and pure image fidelity: Seedance 2.0, if you can get reliable access to it.
The bigger picture
A year ago this comparison would have been Sora against everyone. Last spring Veo had a moment. Then Seedance ate everyone’s benchmarks. Now Omni is in the conversation, and the conversation has shifted.
What changes with Omni isn’t the quality ceiling. It is the ergonomics. The fact that you can edit a video by talking to it, then publish to a billion-user platform without leaving the tool. The other three still ask you to be a producer. Omni lets you be a director.
That isn’t a small distinction. It might be the one that ends up mattering.
