Google unveils Gemini 4 Carbon, Copilot Fusion goes GA, and Genie 4 learns to talk

Gemini 4 Carbon follows Gemini 4 Argon, which shipped two weeks ago, and is tuned for long agentic coding runs and security defense. Google says it ranks first on the Artificial Analysis Intelligence Index v5.0 with 59Vendor claim, and reports 88.2% on the SEC-bench Pro vulnerability-finding benchmarkVendor claim, 54.8% on ExploitGym for exploit validationVendor claim, and 71.0% on CWE-bench for patchingVendor claim. Internally, Google says it kept working for days, backing out of dead ends, to root-cause and fix bugs in large codebases.

For security teams, finding a flaw, checking whether it can be exploited and writing the patch can now run through one model. Pricing matches Argon at $2 per million input tokens and $10 per million output, with cached input 95% off, so teams already on Argon could switch at no extra cost once it opens up. Right now only approved security organizations get in, so most dev teams should track the release date rather than plan a rollout.

Copilot Fusion shows up in the model picker, but it isn't a model. You choose an effort level (Standard, High or Max) and Fusion decides, prompt by prompt, which models to run and how. It has three patterns: Single, where one model answers directly; Cascade, where a cheaper model drafts and a quality check decides whether to escalate; and Critique, where a second model reviews and the first revises. It merges the Project HydraFusion research preview with Copilot's auto model selection.

It cuts the waste of developers picking a premium model for a hard task and forgetting to switch back. GitHub says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code running Opus 5.5 MaxVendor claim, at about half the average API cost per task and roughly 40% less timeVendor claim. Admins can limit which models Fusion may use, or make it the default, which helps control spend across an organization.

Genie 4 is a world model: it generates a 3D scene in real time that you can move through with the keyboard. This version produces ambient sound, music and character dialogue in the same pass as the video. You can type instructions mid-simulation, like changing the weather or adding a person or object. Gemini Omni handles realistic appearance and motion, while other Gemini models keep the scene coherent over time.

The main use is generating endless, varied training environments for robots and for agents that operate screens. Google says early Gemini API and AI Studio partners will get access, which could make this the first time teams building games, simulations or agent test beds can use Genie through an API. Google itself notes that objects and details still drift in longer sessions.