Google’s Gemini 4 Carbon and Genie 4 just changed how AI builds software and worlds

Google is moving at an unrelenting cadence right now in the AI space. Just two weeks after introducing Gemini 4 Argon, the team at Google DeepMind dropped a massive one-two punch across its frontier AI research: Gemini 4 Carbon and Genie 4.

These latest two announcements feel different. Rather than simple conversational upgrades, Google is explicitly targeting persistent autonomous engineering and real-time, steerable world simulation. Between DeepMind’s official reveals for Gemini 4 Carbon and Genie 4, there is a lot to digest. Here is what you need to know about where Google is taking its frontier tech next.

Gemini 4 Carbon tackles multi-day coding and cybersecurity defense

Google is positioning Gemini 4 Carbon as its most capable model yet, specifically engineered for challenging, long-horizon tasks. In internal testing, Google reported that Carbon has been set loose on engineering tasks where it persisted for days through roadblocks to diagnose, root-cause, and patch subtle bugs across massive codebases.

That level of endurance pays off in benchmark scores. Carbon took the #1 spot on Artificial Analysis Intelligence Index v5.0 with a score of 59, while jumping +89 Elo points to reach #4 on WebDev Arena. But the most striking numbers show up in cybersecurity:

• SEC-bench Pro: 88.2% in autonomously discovering real-world vulnerabilities.
• ExploitGym: 54.8% for validating exploitability.
• CWE-bench: 71.0% in turning validated flaws into functional code patches.

Google notes that the Chrome security team saw Carbon generate high-quality patches at more than 10 times the rate of competing frontier models. Importantly, Carbon also completely shut down prompt injection attacks on the Gray Swan benchmark, dropping the attack success rate to a flat 0.0%. Access is rolling out first to cyber defenders and trusted partners through Google’s Fairwind Program before expanding to developers and enterprise surfaces.

Genie 4 brings real-time world simulation with native audio

While Carbon handles the backend engineering heavy lifting, Genie 4 is tackling an entirely different frontier: interactive world generation. DeepMind calls Genie 4 a general world model that delivers real-time, interactive simulation with unprecedented visual realism and native audio generation.

Genie 4 actually leans on Gemini underneath the hood. Gemini Omni anchors the visual appearance and physics of the simulation, while secondary Gemini reasoning models maintain narrative coherence and object permanence over time.

What makes it wild is the steerability. Using simple natural-language prompts on the fly, users can adjust the camera, alter the environment’s weather, or introduce new characters. Genie 4 synthesizes native audio in real time alongside the video – generating ambient background noise, musical scores, and interactive spoken dialogue. DeepMind has already begun rolling Genie 4 out to trusted testers, with plans to bring it to Google AI Ultra subscribers via Project Genie and developers in Google AI Studio.

Whether it is training future autonomous agents or powering a new class of creative workflows, Google’s latest tandem shows a clear vision: models that do not just chat, but actively build, fix, and simulate our digital environments.