Google announced Gemini 4 Carbon on Wednesday, and it is not, for the moment, a model you can use.
The company says Carbon outperformed its own Gemini 4 Argon and OpenAI's GPT-6 Astra on benchmarks that measure reasoning and capability. But Google isn't rolling it out as a general-purpose chatbot upgrade. It's positioning Carbon specifically as a tool for cybersecurity work: finding software vulnerabilities and patching them before attackers do. That's a real, demonstrated skill for this class of model. Just last week, AI-assisted vulnerability hunting helped surface a set of PHP flaws that affected a large share of the world's websites.
Access, for now, is narrow. Only a select group of cybersecurity researchers can use Gemini 4 Carbon today. Asked by CNET when that will widen, Google didn't give a date, saying only that it's working through the US government's review process for the model. Based on how Google has handled past rollouts, broader access could land within one to two weeks, though that's an inference, not a promise from the company.
Genie 4 adds sound to invented worlds
Google's other release this week is Genie 4, an update to its line of world models. These aren't chatbots. Instead of predicting the next word in a sentence, a world model simulates the physics of a space, real or imagined, so developers can build and test things inside it. The obvious uses are robotics and video game design, where you want an AI system to understand how objects move and collide before you build the real thing.
The headline addition in Genie 4 is audio. Previous versions generated silent environments. Now developers can layer in background noise and dialogue, which matters a lot for anyone trying to prototype a game level or a robot-training simulation that needs to sound like a real place and not a vacuum. Google has opened a waitlist for Genie 4 starting now. Rival labs are chasing the same idea: Fei-Fei Li and Yann LeCun are both building world models at their own startups.
A faster pace, a wearier audience
Two new frontier releases in two weeks is a change of rhythm for Google. Gemini 4 Argon arrived 11 months after its predecessor, a glacial pace by the standards of an industry that ships new models almost weekly. Carbon landing this fast suggests Google is trying to close that gap with OpenAI and others.
Whether the public is still impressed is a separate question. Enthusiasm for new AI releases has cooled since last year, and backlash against the infrastructure behind them, particularly data centers, has grown louder. Concerns about AI triggering some kind of civilizational event are, by most accounts, overstated. Worries about job losses, misinformation and deepfakes are not going anywhere.
For now, Gemini 4 Carbon stays behind a government review process and a short list of security researchers. Genie 4 is open to anyone willing to join a waitlist. Where Carbon goes next depends on how fast that federal review moves, and Google isn't saying.
ADFiled by The AI Desk
Models, assistants and the companies and chips behind them, reported from what was released and what was claimed, with the difference kept clear.
More from this desk →
Be the first to comment
Join the argument. No password, just your email or a passkey.