Google is internally testing several Gemini 4 variants, including Argon, Barium, and Carbon, according to leaked documents, screenshots, and chat logs reported by Business Insider. Carbon, which was recently deployed on Google’s internal Jetski coding platform, has demonstrated notable improvements in programming tasks, with one employee comparing its capabilities to Anthropic’s Opus 5.5 coding model.
While early speculation framed Argon as a speed-focused Flash model, Google clarified in official communications that Argon serves as its core frontier reasoning architecture. Internal naming patterns suggest that Carbon and Barium are checkpoints or updates within the Argon model family rather than standalone model tiers, with Carbon likely to ship eventually under the Argon brand.
Google has not released an official launch date for the Gemini 4 family, but recent updates to the Gemini app and Google AI Studio indicate active preparation. DeepMind employee Vedant Misra pointed to recursive self-improvement on social media, echoing recent statements from OpenAI and Anthropic regarding the use of AI to accelerate model development.
Why it matters
Google is leveraging recursive self-improvement to rapidly iterate on coding models, narrowing the gap with Anthropic’s top-tier Opus 5.5.
Argon is officially defined as Google’s flagship reasoning model, replacing legacy tier branding with role-based designations like Flash, Omni, and Gemma.
Developer platforms like Jetski and Google AI Studio offer early signals for upcoming model performance updates before public API launches.
Source: the-decoder.com



