Google's Imagen 3 Just Broke News Before the Reporters

Here's something wild: an AI image generator just scooped the news cycle.

Matt Wolf and Joe Fier from The Next Wave were testing Google DeepMind's new Imagen 3 model (which everyone's calling "Nano Banana 2" because apparently AI needs better branding). They asked it to create an infographic about the Anthropic-Pentagon standoff. When they pushed it to include "the most up-to-date information as of today," the model came back with a timeline that ended with breaking news—news that had dropped 19 minutes earlier.

Neither host had seen it yet.

"Dude, Nano Banana broke the news for us," Fier said, genuinely stunned. "We just had Nano Banana create an infographic. We pressed it for more research data and it pulled freaking breaking news into an infographic."

That's not just fast. That's a different category of tool.

What Actually Changed

Imagen 3 is Google's newest image generation model, and the headline feature is speed—it's roughly twice as fast as its predecessor, Imagen 2 (aka Nano Banana Pro). Where the previous model took around 30 seconds to generate an image, this one consistently clocks in between 12-17 seconds.

But speed alone doesn't explain what happened with that infographic. The model has "search grounding"—it can research topics before generating images. Ask it to visualize how quantum physics works, and it'll actually go read about quantum physics first, then create the image based on what it found.

In theory, this should make AI-generated infographics more accurate. In practice? The results are... complicated.

Wolf tested this with a hometown reference: an infographic of Petco Park and surrounding San Diego landmarks. On the first attempt, with thinking mode set to "minimal," the model got basic geography wrong. "Balboa Theater is definitely not right next door to Petco Park," Wolf noted. "The Western Metal Building is definitely not on the sort of west ocean side of Petco Park."

So they ran it again with thinking mode set to "high." The second version had more detail—dimensions, stadium capacity, specific features. But the spatial accuracy didn't improve much. "It's not great at drawing super accurate representation of what this would look like," Wolf concluded.

Which raises an interesting tension: the model can pull real-time information from the web, but it can't necessarily represent that information accurately in visual space. It knew about the Anthropic blacklisting before the journalists testing it did, but it also thinks buildings are arranged differently than they actually are.

The Copyright Wildcard

One of the more surprising findings: Imagen 3's guardrails are... loose?

Wolf and Fier specifically tested this. They asked for "Super Mario versus Trolls while at Disneyland." It generated the image—Mario included, Disneyland castle in the background. No rejection.

They tried "Donald Trump shakes hands with Dario Amodei." Again, the model complied. Wolf was genuinely surprised: "I'm actually surprised that it didn't just flat out say like no, I can't do that. Honestly." The previous version would reject similar prompts.

Wolf's hypothesis? "What you tend to find with a lot of these models is they'll release them and they have like way less guardrails and then they start to see what people use them for and then they start to like bake the guardrails back in."

If that pattern holds, these tests might not work in a few weeks. But right now, the model is apparently willing to generate copyrighted characters and public figures without much pushback.

The 4K Question

Imagen 3 technically supports 4K resolution, but accessing it requires some navigation. In the standard Gemini interface, Wolf couldn't get images wider than about 2,700 pixels.

In Google AI Studio (which requires a paid subscription), he could select 4K as a resolution option. The output was actually larger than 4K: 5,632 x 3,072 pixels. So if you need high-resolution outputs, AI Studio is the route—but you're paying for it.

The free tier in Gemini will give you decent resolution for most use cases, but not production-quality print work.

The Actual Workflow People Might Use

Wolf's approach to using Imagen 3 is instructive. He doesn't ask it to generate complete YouTube thumbnails ("I've never had amazing luck getting Nano Banana to generate YouTube thumbnails that I'm like really impressed with"). Instead, he uses it to generate assets that he then composes in Canva.

His process: ask Imagen 3 to create a stylized version of his headshot—say, surrounded by bananas. Download that image. Import it into Canva. Add text with a thick outline for visibility. Done.

That's probably the more realistic use case for most people: not replacing design tools, but generating specific visual elements that would be tedious to create or expensive to commission. The model is fast enough now that iteration doesn't feel painful.

What We're Actually Looking At

The Anthropic infographic moment is the thing that sticks with me. Not because the model got everything perfect (it duplicated some timeline entries, the initial version was surface-level). But because it was current in a way that text-based models often aren't.

Most AI models are trained on data with a cutoff date. They're looking backward. Search-grounded image generation is different—it's pulling from the live web, then synthesizing that information into a visual format.

That capability is... I don't know, it feels significant? An infographic that updates itself based on breaking news isn't quite the same as a chatbot that can search. It's doing research and visual synthesis and staying current, all in one tool.

The accuracy issues are real—you absolutely need to fact-check anything it produces before using it. But the fact that it can pull information that's minutes old and turn it into a coherent visual? That's a different kind of tool than we had a year ago.

Google's made this free in Gemini, which means a lot of people are about to start using it. Whether the guardrails stay loose or tighten up in response to how people actually use it—that's the part I'm watching.

— Yuki Okonkwo, AI & Machine Learning Correspondent