Gemini Omni Explained | Any-to-Any AI Video, Conversational Editing & the End of Sora's Moat

Añadido: 2026-05-25

[00:00:00]You've probably been bouncing between five different AI tools just to make a single video. One for the script, one for the voice, one for the visuals, another for editing, and somehow it still looks like AI made it. Trust me, I burned weeks of my life doing exactly that.

[00:00:15]But last week at Google I/O, Google quietly dropped something that made every single one of those tools feel kind of pointless, and I'm not exaggerating. What I'm about to show you is the biggest shift in AI video we've seen since Sora. Welcome back to bitbias.ai, where we do the research so you don't have to. Join our community of AI enthusiasts with our free weekly newsletter. Click the link in the description below to subscribe. You will get the key AI news, tools, and learning resources to stay ahead.

[00:00:45]So in this video, I'll break down exactly what Google's new Gemini Omni model is, what it can actually do today, and how it stacks up against Sora, GPT-4o, and Claude. So by the end, you'll know whether it's worth paying for or whether you should wait. First up, the part that nobody's talking about. This isn't just another video generator, it's something completely different. Let me show you what Gemini Omni actually is. Okay, so here's the simplest way to think about it. Every AI video tool you've used until now works like a relay race.

[00:01:19]You write a prompt, it gets handed to a text model, which hands it to an image model, which hands it to a video model, and every handoff is where things break down. The lip sync goes off, the character's shirt changes color, the lighting jumps. Gemini Omni throws that whole system in the trash. It's what Google calls an any-to-any model, which means a single model that takes in any combination of text, images, audio, or video, and spits out high-quality video on the other side.

[00:01:49]One brain, no handoffs. The first model in the family is called Gemini Omni Flash, and it went live on May 19th, 2026.

[00:01:58]If you're on Google AI plus pro or ultra, you already have it inside the Gemini app, inside Google Flow, and even inside YouTube Shorts. Yeah, YouTube Shorts. We'll come back to that because that one's a really big deal for creators. But here's where it gets interesting. The thing people are losing their minds over isn't the generation, it's the editing, the conversational editing workflow. Imagine this, you generate a video of a violinist playing in a concert hall. Pretty cool, but then you just talk to it.

[00:02:27]Transport the violinist to a beach.

[00:02:30]Done.

[00:02:31]Now make the violin invisible.

[00:02:34]Done.

[00:02:35]Make it snow.

[00:02:37]Done.

[00:02:38]And every one of those edits builds on the last one.

[00:02:41]The scene stays continuous, the character stays consistent, the lighting evolves naturally. This is what Google's calling conversational editing, and the reason it works is exactly what I mentioned earlier.

[00:02:53]There's no handoff between models. The same brain that generated the first frame is the brain doing the edit, so it remembers everything. The lake you added two prompts ago is still there. The character you swapped in still looks the same.

[00:03:06]If you've ever tried to do this in Runway or Pika, you know what a nightmare it is. You get one good clip, you try to extend it, and suddenly your main character has three arms and a slightly different face. Omni mostly fixes that, and I say mostly because we'll get to the limitations in a bit.

[00:03:22]Drop a comment if you've hit that exact frustration before.

[00:03:25]I want to see how many of you have rage quit a video project because of it. How it's built. Now let's get a little nerdy for a second because the architecture here is actually important, and don't worry, I'll keep it human. Omni Flash is a transformer model, just like the LLMs you already know. But the trick is that it was built from day one to handle text tokens, image patches, audio frames, and video frames in the same shared model.

[00:03:49]Most older systems would learn each modality separately and try to glue them together. Google trained Omni on all of them at once. To pull that off, Google reportedly used their new TPU V8 cluster running on Jacks and Pathways.

[00:04:02]And Sundar Pichai mentioned they've now scaled training to clusters of over a million TPU chips.

[00:04:08]That's not a typo.

[00:04:10]A million, which is a useful thing to remember the next time someone tells you the AI race is slowing down. It really, really isn't. Google didn't release the exact parameter count, which is honestly pretty standard for frontier models these days. But the practical takeaway for you is this. Omni Flash is small enough and efficient enough that generation feels nearly interactive. Not instant, but not the 5-minute waits you might be used to from earlier video models. Coming up, the part that matters most for anyone actually trying to use this. What it costs, what you can do today, and where Google is pulling some sneaky moves. Access, pricing, and the catch. Okay, here's the breakdown.

[00:04:49]To use Omni right now, the basic tier, AI Premium, is $7.99 per month. And that gets you in. AI Premium sits in the middle.

[00:04:59]And then there's the new one Google just rolled out. AI Ultra at $100 a month, which gives you priority access, higher usage caps, and bundled access to Omni plus Google's new agent tools like Anti-Gravity 2.0. Now, here's the catch.

[00:05:14]And this is the part Google is being weirdly quiet about. There is no public API yet. Not on Vertex, not anywhere.

[00:05:22]Google said it's coming in the coming weeks, which in Google speak could mean anywhere from next Tuesday to never. So, if you're a developer trying to build a product on top of Omni, you're stuck waiting. Right now, Omni is a consumer feature dressed up like a frontier model.

[00:05:36]That said, what you can do today is pretty wild.

[00:05:40]Inside the Gemini app, you can generate and edit short videos with text and reference inputs. Inside Google Flow, which is their new web-based AI creative suite, you've got a full canvas for building scenes. And inside YouTube Shorts remix, you can take an existing short and reimagine it. Change the setting, swap a character, add effects, whatever.

[00:06:01]If you're a creator, that last one is the sleeper feature. Imagine grabbing a trending short, remixing it with your own twist using Omni, and posting it in 5 minutes. That's a content workflow that didn't exist a week ago. How it compares to Sora, GPT-4o, and Claude.

[00:06:17]All right, let's do the comparison everyone's been asking for.

[00:06:20]And I'll be honest with you, the picture is messier than the headlines suggest.

[00:06:24]OpenAI's Sora came out in late 2024 and was the model that made everyone realize AI video was real. It still produces gorgeous footage, and it can go up to about 60 seconds.

[00:06:36]But Sora is text in, video out. You can't drop in an image, you can't drop in audio, and you definitely can't have a conversation with it the way you can with Omni. It's a one-shot machine.

[00:06:47]Beautiful, but rigid. GPT-4o.

[00:06:50]OpenAI's so-called Omni model from 2024 was actually the first one to use that Omni branding. But here's the thing nobody really emphasizes. GPT-4o never generated video. It does text, image, audio, and code. That's it. And it's mostly been folded into newer GPT versions at this point. Claude 4 Opus from Anthropic is the reasoning king.

[00:07:15]It's incredible for analysis, for long context, for writing. It can look at images, but video? Not its game. So if you actually line them up, Gemini Omni Flash is in a category of one right now.

[00:07:27]It's the only model that takes any input, text, image, audio, video, and gives you a generated video back with conversational editing on top. That's the moat. But, and this is important, moats in AI last about 3 months. Sora 2 is rumored. Open AI has been hiring video researchers like crazy and Anthropic just keeps quietly shipping.

[00:07:50]So, enjoy this gap while it lasts. Quick question for the comments. Which lab do you think catches up first? Open AI, Anthropic, or somebody we're not even watching?

[00:08:01]I'm genuinely curious what you all think. The real limitations nobody's mentioning. Okay, time for the part of the video where I tell you what Google's marketing team would rather I didn't.

[00:08:10]First, the API thing. I already mentioned it, but it's worth repeating.

[00:08:14]If you're a developer or you run a business, you can't build on Omni yet.

[00:08:18]So, all those startup demos you're seeing on Twitter, most of them are clever wrappers around the Gemini app, not real integrations. Second, compute cost at scale. Video generation is brutally expensive. Even though Google claims Omni flash is way faster and more efficient than other frontier models, generating a few seconds of high resolution video still chews through serious compute. When the API does drop, the per token pricing for video could be eye-watering. We just don't know yet.

[00:08:44]Third, the watermark conversation.

[00:08:47]Google built in something called SynthID, which is an invisible watermark on every generated image and video. They also rolled out content credentials, so Chrome and search can flag AI edited content. That's genuinely good for transparency.

[00:09:02]But, it also means everything you make is traceable to Google servers, which raises real privacy questions if you're feeding it sensitive footage. Fourth, and this is the one I'd actually watch out for, physics and edge cases.

[00:09:15]Google says they've improved physics simulation and the demos do look better than older models, but Google didn't publish any standard benchmarks.

[00:09:23]So, we're taking their word for it.

[00:09:25]Expect weird hallucinations in complex scenes, especially anything involving multiple people interacting, hands, or fluid motion. And finally, lock in. The more you build your workflow around Omni, the harder it gets to switch later.

[00:09:39]That's not unique to Google. It's the AI game in general. But it's worth saying out loud. What comes next and what you should do?

[00:09:46]So, where does this go from here? Google was pretty open about the road map. Omni right now is focused on video output.

[00:09:52]But they confirmed that future Omni models will generate images and text natively, too. That's the full vision of any modality to any modality. One model that handles everything. We're also going to see Omni show up inside more Google products.

[00:10:06]Expected in search, in workspace, in maps, eventually probably in Android XR glasses.

[00:10:13]And on the agent side, Google's new Gemini Spark, which is basically a 24/7 cloud assistant, is going to lean heavily on Omni for anything that involves visual output. So, what should you do this week? Three things. One, if you've got a Google AI subscription, open the Gemini app and actually try Omni.

[00:10:31]Don't just watch demos.

[00:10:33]The fastest way to understand what this thing can do is to break it yourself.

[00:10:37]Try weird prompts. Try editing.

[00:10:40]See where it falls apart. Two, if you're a creator on YouTube, get into the Shorts remix feature. There's a window right now where most creators don't know this exists.

[00:10:49]And that's where opportunity lives.

[00:10:51]Three, keep an eye on Google AI Studio and the DeepMind blog over the next few weeks for the API drop. The moment that goes live, the entire AI app ecosystem is going to shift. All right, that's Gemini Omni. The short version, it's a genuinely new kind of model. It's the only thing in its category right now.

[00:11:09]And it's going to push every other lab to ship faster. The long version is everything I just showed you. If this broke down the announcement in a way that actually made sense, hit the like button. It really does help the channel.

[00:11:21]And subscribe if you want me to keep covering every major AI launch the same way. Drop your honest take in the comments. Are you switching to Omni, sticking with Sora, or waiting it out? I read all of them.

[00:11:32]I'll see you in the next one.