Google’s Gemini 4 Carbon test raises a question about Argon

Argon’s place in the lineup

Google describes its model roles this way: “Argon is for advanced reasoning, Flash is for speed and high volumes of requests, Omni is for media generation, and Gemma is for lightweight tasks on devices with open weights.”

That puts Argon in the company’s top reasoning tier, alongside Anthropic’s Opus and OpenAI’s Astra. Carbon and Barium, however, may be checkpoints or updates within the Argon family rather than separate model classes. One Google employee called Carbon “the next Gemini Pro model,” which could mean it will ship under the Argon name. Internal documents offer a precedent: the Argon model already introduced publicly was previously called “Barium-B.”

Whether Carbon becomes an Argon update or a distinct model is still unclear. Frequent releases resembling Google’s Flash updates may indicate the company is getting better at using AI to build new models. Google DeepMind employee Vedant Misra responded to a Business Insider post on X by discussing recursive self-improvement: using AI to develop better AI. OpenAI and Anthropic have reported similar progress.

Google Deepmind employee Vedant Misra points to recursive self-improvement as a driver behind the rapid model iteration. | via X

Source: the-decoder.com

Signs of a Gemini 4 launch

Google has not announced a release date for Gemini 4, but it is already changing its apps. The Gemini app has a new “Automatic” mode for third-series models and a reasoning-intensity setting that ranges from low to high. Google AI Studio has shown an “Ultra” mode, described as offering expanded capabilities and tools. Google’s Logan Kilpatrick confirmed that the team is working to make the most of Argon’s capabilities.

I think the Carbon-versus-Opus comparison is less useful than the naming question. If Carbon is simply a new Argon checkpoint, the meaningful change may be how quickly Google can improve its reasoning model—not a new model category. The announcement is quiet about the evidence behind the programming comparison, leaving the headline claim harder to assess than Google’s product signals.