Good morning, AI filmmakers.

Google I/O 2026 hit Tuesday. In one keynote: a frontier model, a video builder, a world simulator that ingests Street View, and an agent layer tying it together — software that doesn't just answer, it does the work.

The agentic shift everyone's been promising — Google just declared they're shipping it first. Pichai: "we're firmly in our agentic Gemini era."

What matters for filmmakers isn't what Google said. It's what it implies. Last week I wrote about open-source AI video tools finally fitting into a workable home pipeline. This week Google moved the other way — put the whole creator stack inside one product line and said: come live here.

Two opposite pulls hitting the same year. You'll feel both.

In today’s AIography:

Read time: About 5 minutes

WHAT I’M THINKING ABOUT

Four announcements matter for working filmmakers: Gemini 3.5 Flash, Gemini 3.5 Pro, Gemini Omni Flash, and Genie + Street View.

3.5 Flash is available today. According to Koray Kavukcuoglu, who runs DeepMind's technology side, "3.5 Flash offers an incredible combination of quality and low latency. It outperforms our latest frontier model, 3.1 Pro, on nearly all the benchmarks." The smaller, cheaper variant beats the larger model Google was shipping six months ago. If it holds in practice, that's the cost-curve story behind everything else.

3.5 Pro rolls out next month. Omni Flash, the video builder, is available now to Google AI subscribers through Flow and Flow Music. Genie + Street View is available to select Ultra-tier subscribers in the U.S., with global rollout in following weeks.

The I/O announcement was a stack announcement, but not all of the stack is on the shelf yet. Some pieces are. Some are still measured in "next month" or "select users." Knowing what's actually shippable from each I/O keynote is what separates a credible vendor read from a press-release recap.

Omni Flash is Google's answer to Sora, Runway Gen-4, Kling, and Veo. Google's framing: Omni is "a world model AI that can understand and simulate the world." Translation: the model is supposed to hold environment state across multiple shots so the same scene stays coherent between cuts. That's the part that interests narrative filmmakers.

Demis Hassabis called Omni "our new model that can create anything from any input." Google's blog post doesn't yet specify resolution, shot length, or failure modes. The product is available today through Flow, and the working creator's question is the one demo reels cannot answer: how does it compare on a shot you actually need?

Two things to watch. Google has bet on "world simulator" as a marketing frame. The first benchmark people will run is "does the same character stay the same across multiple shots." If it does, Omni separates from Sora and Veo on a real axis. If it doesn't, the framing collapses fast.

Second: Omni comes paired with a music generation layer (Flow Music) on the same product surface. The audio side of the AI video stack has been the missing piece for short-form creators all year. Google is now offering both in one subscription.

This is the announcement that got the most filmmaker-facing trade press, and the one that needs the most careful reading.

The headline: Genie now ingests Street View imagery and turns real-world locations into interactive simulations you can move through. Weather can be adjusted. Viewpoints shift between humans, robots, and other agents. The immediate read is "virtual location scouting just got easy."

That read is incorrect. Or more fairly, six to 12 months early. Jack Parker-Holder, the DeepMind researcher who built Genie, told TechCrunch on the record: "It's maybe six to 12 months behind video in terms of accuracy and quality."

The TechCrunch piece also notes that output is "video game quality rather than photorealistic," that the model "is not yet physics-aware; characters pass through obstacles," and that models "cannot yet create faithful street reconstructions." Those are not small caveats. The difference between "I can scout this for my next project" and "this is a tech demo" runs right through that list.

Where Genie matters in its current form is previsualization, not location plates. If you direct narrative work and need to block a scene before booking the real location, the video-game-grade output is useful. If you're trying to generate an establishing shot to drop into a finished cut, you are not there yet by Google's own admission.

The pattern to watch is the same one that played out with text-to-video two years ago. First demos look great until you try to do real work. Then the seams show. Then six to 12 months pass and they don't. Plan accordingly.

The most interesting thing in the 3.5 announcement is the workflow shape Google is selling. Tulsee Doshi, who runs product at DeepMind, put it this way: "3.5 Pro becomes your orchestrator, your planner, and then it actually can leverage Flash to be the various sub-agents."

A subagent is a smaller faster instance of the model that takes a piece of the work. Pro plans the project. Flash does the pieces in parallel. You issue one instruction, the system breaks it into ten tasks, runs them concurrently on cheap fast Flash instances, and pulls the results back together.

For creative pipelines this is the cost story. If an AI-driven storyboard-to-shot pipeline used to cost X dollars per minute, the Pro-orchestrates-Flash architecture moves the math. Most compute happens on the cheap variant. Only planning happens on the expensive one. Multiplied across an episode or a season, that changes what's economically possible for AI-assisted production.

The use cases Google highlighted are mostly enterprise (banks, data teams). Whether a multi-step creative pipeline — concept to storyboard to shot list to image to video to music — actually runs end-to-end on this stack is the test working filmmakers will pressure-check over the next 30 days.

While Google ran I/O, OpenAI shipped a tool that checks whether an image was generated by an OpenAI model. The tool reads two signals: C2PA metadata, a standardized origin tag embedded in the file, and a SynthID watermark, a Google-developed invisible pattern baked into AI-generated pixels.

A watermark here is not a visible logo. It's a pattern of pixel changes invisible to the human eye that detection tools read mechanically. Both signals can be stripped by determined adversaries. Together they're better than either alone.

The interesting move is cross-vendor. OpenAI implemented SynthID, a Google standard, in their own products. The watermarking debate has been whether the major labs converge or fragment. This week, OpenAI signaled they're converging.

The tool currently only detects OpenAI-generated images. It doesn't catch Midjourney, Gemini, Flux, or any open-source model. Expansion to other vendors is described as "aspirational, not guaranteed." A working filmmaker cannot yet use it as a universal AI-image detector.

What's coming is the world where you can. Clients and platforms are starting to ask for AI-content disclosure. The watermarking infrastructure being built this week makes those disclosures mechanically verifiable instead of trust-based. Worth knowing what shipped.

ACTION ITEM

If you already pay for Google AI, open Flow and run a five-shot continuity test through Omni Flash. Pick a character you've been working with. Generate the same character in five different setups. Look at whether the face, the wardrobe, and the lighting hold. That tells you whether the "world simulator" framing is real or marketing.

If you don't pay for Google AI, the test is cheaper. Spend 30 minutes reading the Decrypt piece on Omni Flash and the TechCrunch piece on Genie. Note which specific capabilities each source actually claims, versus the bullet-point summaries the trade press is running. The gap between headlines and source material this week is wider than usual. Vocabulary for the next six months of client conversations: agentic Gemini, Omni Flash, Genie world model, SynthID watermarking, C2PA metadata.

For everyone else, block out time to revisit Omni and Genie in 60 days. The honest read of either cannot be made from launch-week press.

ONE MORE THING…

Corridor Crew rebuilt one of the great in-camera tricks from Lord of the Rings this week, a 25-year-old shot Peter Jackson's team solved with motorized rigs, forced-perspective set construction, and live optical work instead of doing it in post. Watching Corridor walk through it is a useful counterpoint to a week of AI-stack announcements. The shot held up because the framing solved the problem before the camera rolled.

Google announced a world simulator on Tuesday. The simulator can't yet hold physical state. The shot Jackson's team built can. The seams in AI tooling are still in roughly the places craft-driven filmmakers have always solved with planning, blocking, and patience.

SHORT TAKES

ESSENTIAL TOOLS

Blackmagic released Fusion Studio 21.0 in beta this month. For compositing and visual effects work inside the DaVinci Resolve ecosystem, this is the release cycle where it's worth opening a test project and seeing what changed. Beta cycles are the right time to find the workflow breakers before they ship into the production version that lands in your client work.

If you composite AI-generated elements into live-action plates, Fusion remains the working filmmaker's tool of choice for the node-based work that AI-heavy projects increasingly require.

Check out the Alpha version of our AI Tools Database. We will be adding to it on a regular basis.

Got a tip about a great new tool? Send it along to us at: [email protected]

FINAL THOUGHTS

Last week the open-source AI video stack quietly crossed the bar where it can be tested seriously against the paid cloud stack. This week Google announced a counter-move. One product, one subscription, the whole creator-facing pipeline. Two vectors. One year. Both real.

The honest read on what to do about it is the routing question stays open per shot. Some shots run cheapest and best on the open home stack. Some shots run cheapest and best inside Google's new integrated pipeline. Some shots still run best on Runway or Kling. The right pipeline this year is the one that asks the routing question for every project instead of locking into a single vendor on faith.

Google's I/O was a stack-shipping announcement, but the stack isn't all on the shelf yet. Some of it is six to 12 months out by Google's own admission. The trade press will spend the next 30 days inflating capabilities the source documents don't actually claim. The way to read it is to wait for practitioner reviews and revise your routing decisions when the data lands instead of the headlines.

Stay sharp. Keep creating.

— Larry

If you have specific feedback or anything interesting you’d like to share, please let us know by replying to this email.

AIography may earn a commission for products purchased through some links in this newsletter. This doesn't affect our editorial independence or influence our recommendations—we're just keeping the AI lights on!