Google Tests Gemini 4 Carbon Variant to Boost Coding

Internal communications reveal that Google is running trials for several Gemini 4 variants, including codenames Argon, Barium, and Carbon. Carbon was recently rolled out to Google employees on its internal development software, Jetski, where early testers noted significant gains in software engineering tasks compared to Argon. One worker noted that Carbon's programming performance rivals Anthropic's unreleased Opus 5.5, whereas earlier Argon iterations were closer to Opus 5. Argon itself is categorized as Google's primary frontier reasoning model, alongside Flash for speed, Omni for media, and Gemma for edge devices.

Carbon and Barium appear to be developmental checkpoints within the broader Argon system rather than standalone tiers, with internal documents indicating Argon was once labeled Barium-B. Responding to leaks about the rapid pace of model updates, Google DeepMind team member Vedant Misra alluded on X to recursive self-improvement—the concept of using existing artificial intelligence systems to engineer superior successor models. Competing labs such as OpenAI, with its Astra framework, and Anthropic are reportedly leveraging similar automated feedback loops to accelerate their model pipelines.

Ahead of an official launch date, Google has begun modifying its public tools to support these upcoming capabilities. The consumer Gemini interface recently added an Automatic mode for legacy 3-series models along with controls to adjust reasoning depth between low and high settings. Meanwhile, Google AI Studio has added an Ultra preset designed for complex tool usage, and Google's Logan Kilpatrick confirmed active efforts to maximize Argon's output. For software developers, these leaks signal an impending jump in automated code generation and reasoning tools integrated directly into Google's developer ecosystem.