I've Been Tracking Gemini Omni Leaks for Six Weeks. Here's What I Think Google Actually Launches on May 19

I have a confession to make. For the last six weeks, I've been watching AI video leaks the way other people watch sports.

Every morning, before coffee, I check four sources: 9to5Google, Chrome Unboxed, TestingCatalog, and a handful of developer Discords I probably shouldn't be in. I'm looking for screenshots. For metadata strings. For the specific way a leaked demo's chalk hits a chalkboard in a model that doesn't officially exist yet.

This is not normal hobby behavior. I'm aware of that.

But Google I/O 2026 is six days away as I write this, and something significant is about to drop. Something that I think most people — even people who follow this stuff — have not quite understood the size of yet.

The model is called Gemini Omni, or possibly Veo 4. Even that ambiguity is part of the story.

## The Day I Realized This Was Different

It was a Tuesday in early April. A leaked demo started circulating in a Telegram channel I follow. The clip was twelve seconds long.

A professor was writing a trigonometric proof on a green chalkboard. Just that. Chalk dust caught in the light. A hand moved across the surface. The proof resolved correctly. The hand pulled away. End of clip.

I've watched a lot of AI-generated video over the past two years. I've watched Sora's surreal, oversaturated dreams. I've watched Runway hallucinate fingers into octopi. I've watched Seedance render almost-real food that always seemed to be made of plastic if you looked too long.

This was not that.

The chalkboard demo wasn't impressive because it was photorealistic. It was impressive because the chalk *behaved like chalk.* Each stroke left the right kind of residue. The proof was written, not generated as a finished image — the model knew the difference between drawing a 7 and being mid-stroke through a 7. The professor's hand didn't smear what was already there.

That sounds like a small thing. It's not. It means the model has a kind of temporal understanding that previous video AIs simply do not have. It understands that the world has a state, that the state changes over time, and that what happens at second three depends on what happened at second one in a causally consistent way.

That's the kind of capability that doesn't just make better videos. It makes a different category of tool entirely.

## The Sora Retreat That Almost Made Me Stop Watching

I was honestly about to write this off as another impressive demo that wouldn't ship the way the demos suggest. AI video has a long history of leaks that turn out to be ten percent of what eventually arrives. Then I remembered what happened three weeks earlier.

On April 29, OpenAI quietly shut down the consumer-facing Sora 2 app. They kept the underlying model — it's still available through the API — but the consumer app, the place where ordinary people could actually generate videos with it, was gone. The framing was the usual corporate language about "refining the experience" and "distribution focus." The substance was a retreat.

I sat with that news for a few days. OpenAI doesn't retreat from consumer markets. They invented several of them. The fact that they walked away from consumer video specifically — while keeping the technical capability alive on the paid API tier — told me something I hadn't fully internalized: AI video generation is genuinely expensive in a way that breaks the unit economics of free or near-free consumer access.

So when the Gemini Omni leak hit eleven days later, the timing stopped feeling like coincidence. It read like Google making the opposite bet. OpenAI looked at the math and said: not yet. Google looked at the same math, factored in their TPU buildout and the established Gemini user base they can amortize the cost across, and said: yes, now.

One of them is wrong about this. We're about to find out which.

## The Two Realities I Keep Switching Between

Here is the part where I'm supposed to tell you this is great, hand you a list of five ways AI video will transform your business, and end with a call to action.

I'm not going to do that, because I keep flipping between two genuine feelings about this, and I think you should know about both.

**Reality one:** This is incredible. If the leaks are even half right, what's about to launch will let one person produce, in an afternoon, work that used to require a small production company. Voiceover, soundtrack, on-screen text, scene transitions — all from a prompt. All editable through a conversation, not a regeneration. I've been writing about AI tools for years and I have not been this excited about a single release since GPT-4. The creator economy gets meaningfully larger because the floor of who can make professional-feeling video gets meaningfully lower.

**Reality two:** I do not know what this means for the people who already do this for a living. I don't know what it means for the friend of mine who left her advertising job last year to freelance as a video editor. I don't know what it means for the wedding videographers, the corporate explainer shops, the small post-production houses that have built careers on the assumption that this work requires human craft.

I think a lot of those careers are about to change in ways that the launch keynote is not going to address.

Both of these things are true at the same time, and the discourse online wants me to pick one and shout about it. I'm not going to.

## What the Compute Bill Tells Us

There is one specific leaked screenshot that has been bothering me for weeks.

A Gemini AI Pro subscriber posted a screenshot of their daily usage quota. Two short video generations had consumed approximately 86 percent of the daily allowance.

Eighty-six percent.

For a subscriber paying around twenty dollars a month, that means roughly two videos per day at the consumer tier. That's not a productivity tool. That's a curiosity allowance.

What this tells me is that whatever Google is about to launch is *expensive to run.* Estimates from the infrastructure community suggest twelve to twenty times the compute cost of current production video models. The price will come down — it always does — but the early launch period is going to be defined more by rate limits and waitlists than by the kind of "use it for everything" experience the demos suggest.

It's not unique to Google, either. ByteDance's Seedance 2.0, which sits at the top of most public video benchmarks right now, has tiered its access aggressively. Alibaba's Wan 2.7 — which ships the most comprehensive feature set in the field, including audio-synced 1080p output — caps free generation at single-digit clips per day. Kuaishou's Kling V3.0 prices its highest tier above premium ChatGPT Plus. Everyone is dealing with the same compute reality. The question for Gemini Omni isn't whether the constraints exist — it's whether Google's infrastructure advantage lets them push consumer pricing closer to "barely sustainable" than anyone else can manage.

The honest version of the story is: the capability is real, the access is going to be rationed, and the people who get to use it heavily in the first six months are going to develop a meaningful skill advantage over everyone else who's waiting in line.

## What I Will Actually Be Watching For on May 19

The keynote will happen. The demos will be polished. There will be moments designed to go viral.

I'm going to be watching for five things underneath all of that:

**The name.** If Google calls it Veo 4, that signals continuity with their existing enterprise positioning. If they call it Gemini Omni, that signals a consumer-first strategy with everything else as secondary. The branding choice is the strategy.

**The price.** Aggressive pricing means Google wants to crush specialized tools out of existence. Premium pricing means Google sees this as a complement to existing workflows. Both are defensible. They imply very different next eighteen months.

**The API timing.** Same-day Vertex AI availability means Google is confident in the model. Delayed enterprise access means they want more validation. The gap between consumer launch and developer access is itself a signal.

**The quotas.** If the daily generation limits feel generous on day one, demand is being met. If they feel painful, the launch is more about strategic positioning than market penetration.

**The benchmark numbers.** Gemini Omni will be compared, on launch day and every day after, against Veo 3.1, Sora 2's remaining API, ByteDance's Seedance 2.0, Alibaba's Wan 2.7, and Kuaishou's Kling V3.0. If Google ships without a clear lead on at least two of the standard axes — physical realism, prompt adherence, audio sync, camera control — the launch narrative becomes "workflow advantage" rather than "capability leadership." Those are very different positioning stories, and the long-term consequences for who wins this market are larger than the keynote will admit.

These five things will tell me more about what's actually happening than any of the keynote speeches will.

## Why I Built a Tracker for All of This

I'm going to be transparent about something: I've been collecting all of this information in one place because I needed it for myself first. The leaks are scattered across dozens of accounts. The technical details disappear into thread replies. The actual story of what Google is building has been hiding in plain sight, just spread across too many surfaces to follow.

Other people I've talked to have the same problem. So I made the tracker public.

If you want to follow what actually drops on May 19 — not the keynote version, but the developer documentation, the rate limit reality, the leaked benchmarks against Sora 2 and Seedance — you can. I'll be updating in real time during the keynote and the week after.

That's not a sales pitch. That's just me telling you where I'll be the night of May 19, refreshing developer subreddits at 11 PM Pacific time, taking notes, trying to figure out what just happened.

## A Final, Honest Thought

I've been writing about AI for long enough now that I've learned to distrust my own initial reactions. The first wave of any major launch is always either too hyped or too dismissive, and the truth is usually in the boring middle that nobody clicks on.

What I think will happen on May 19 is this: the launch will be more interesting than the launch-day hot takes suggest, less revolutionary than the most enthusiastic adopters claim, and meaningfully different from what existed two weeks before in ways that take six to twelve months to fully understand.

Some workflows will get transformed in obvious ways. A larger set will change in subtle ways that don't show up in the first month of usage data. A surprising number of jobs that nobody currently associates with AI video will turn out to be partially automatable in ways nobody saw coming.

The interesting part of this launch isn't what Google announces. It's what we collectively make of it in the months after, in studios and offices and bedrooms and weddings and YouTube channels and corporate training departments, when the actual humans who do video work for a living sit down with a new tool and start figuring out what it changes.

I'm going to be paying attention. If you want to pay attention with me, you know where to find the notes.

Six days. Refresh.