Google Ships Gemini 4 Carbon and Genie 4, Its Second Frontier Release in Two Weeks

Google announced two new AI models on Wednesday: Gemini 4 Carbon, its most advanced large language model to date, and Genie 4, a world model that can now generate audio. The releases come just two weeks after Google's last frontier model launch, a pace shift for a company that spent most of 2025 watching rivals ship.

What did Google actually release?

Two models, built for very different jobs.

Gemini 4 Carbon is a large language model — the same basic technology behind chatbots. Google says it's the most advanced model the company has released. On benchmarking tests designed to assess a model's overall capabilities, it surpassed both Google's own Gemini 4 Argon and OpenAI's GPT-6 Astra, the company claims.

Genie 4 is something else entirely: a world model. Where a chatbot predicts the next word in a sentence, a world model simulates real and imagined physical worlds, grounded in the physics of a digital universe. Genie 4 adds a new capability to the lineup — audio. Developers can now add background noise and dialogue to their AI-generated worlds.

The stakes are practical. Google pitches world models as useful for robotics and video game design, two industries that need machines capable of reasoning about how objects move and interact. The field has serious pedigree: Fei-Fei Li and Yann LeCun, two of the biggest names in AI research, are building world models at their respective startups.

Why is Google pitching a frontier model at cybersecurity?

Because vulnerability detection has turned out to be one of the most consequential things these models can do. Google says Gemini 4 Carbon is good at detecting and patching vulnerabilities — and that skill has already proven itself in the real world. Just last week, AI-assisted vulnerability research surfaced PHP flaws affecting a majority of the world's websites.

That's the context for the restricted rollout. Neither Gemini 4 Carbon nor Genie 4 is available to the general public yet. Only select cybersecurity experts have access to Gemini 4 Carbon right now. Google declined to answer a question from CNET about when access will open up, but if the company follows the pattern of previous releases, it may be as soon as one to two weeks from now.

There's a regulatory dimension too. Google says it is actively engaging with the US government's review process — a signal that frontier releases now pass through a policy gate, not just an engineering one. Anyone who wants Genie 4 can join the waitlist now.

How far behind was Google, really?

The framing of an "AI race" needs some unpacking, because the numbers cut both ways.

Google ended 2025 as a major player in the AI industry. But it hadn't kept up with its peers on raw release cadence. Gemini 4 Argon, the company's previous frontier model, was its first new frontier release in 11 months. That's a long stretch in an industry that sees new models seemingly every week.

Two releases in two weeks suggests Google is done waiting. Benchmark wins over GPT-6 Astra give the company a marketing claim; the cybersecurity positioning gives it a use case that regulators and enterprise buyers can evaluate concretely.

The takeaway: Google has moved from a slow release cycle to a fortnightly one, and it's doing so with a model aimed at a specific, high-value professional audience.

Why does public mood matter for this launch?

Because the market these models are landing in looks nothing like the market of a year ago.

Public appetite for AI has changed a lot since 2025, and not in the industry's favor. Fears around AI have risen. So has backlash against AI infrastructure projects such as data centers — the physical footprint that makes models like Gemini 4 Carbon possible in the first place.

Not all of these fears carry equal weight. As CNET notes, fears about AI contributing to human extinction are often overblown — or, perversely, actually hand tech companies more power by framing regulation around distant catastrophes. But other concerns are concrete and current: mass job losses, misinformation, and deepfakes are all live problems, not hypothetical ones.

That tension shapes what a launch like this one means. A model that finds and patches vulnerabilities is an easy story to tell a skeptical public. A model that simulates convincing worlds with sound and dialogue is a harder one, given how those same capabilities overlap with deepfake and misinformation risks.

What happens next?

Watch the access timeline. If Google holds to its historical pattern, Gemini 4 Carbon could reach a broader audience within one to two weeks. The US government review process is the wild card — Google says it's "actively engaging" with it, and that engagement could stretch or shorten the wait.

The Genie 4 waitlist is open now, which means developers will start testing the audio capability — and its limits — soon. Between benchmark claims, a restricted rollout, and a federal review, Google's fortnightly pace shows no sign of slowing; whether public trust keeps up is the open question.