Google is testing a new Gemini 4 artificial intelligence model codenamed Carbon, which internal testers report matches the coding capabilities of Anthropic's flagship Opus 5.5 model. Unreleased Carbon and early variant Barium-B belong to the Gemini 4 family announced in September 2026, where Argon serves as the core frontier reasoning tier. Google DeepMind researcher Vedant Misra cited recursive self-improvement as a driver of rapid model development as Google preps app updates and workplace agents to counter competitors like OpenAI.
Gemini 4 Carbon testing and capabilities
- ▪Google internal discussions referred to the Carbon variant as 'Gemini pro next model,' indicating it could be deployed as a future checkpoint or update to Argon
- ▪Google employees began internally testing an unreleased Gemini 4 model variant codenamed Carbon on the company's internal coding platform Jetski in recent days
- ▪Google internal testers reported that the Gemini 4 Carbon model demonstrated coding capabilities comparable to Anthropic's flagship Opus 5.5 model for long-term agentic programming
Gemini 4 Argon release and strategy
- ▪Google announced its new generation of Gemini 4 models in September 2026, establishing Argon as its primary frontier reasoning model
- ▪Internal Google documents revealed testing for a series of Gemini 4 models named Argon, Barium, and Carbon, with the variant previously called Barium-B selected for public release under the Argon name
- ▪Google designated specific roles across its model portfolio, pairing Argon for frontier reasoning alongside Flash for speed, Omni for generative media, and Gemma for lightweight edge tasks
Internal reception of Argon
- ▪Google employees evaluating early internal builds of the Gemini 4 Argon model noted that its performance on certain coding tasks lagged behind Anthropic's older Claude Opus 5 release
- ▪Google staff gave generally positive overall feedback during internal testing of early Gemini 4 Argon model versions
Gemini application and studio updates
- ▪In preparation for the Gemini 4 release, Google updated the Gemini application with an Automatic mode and adjustable reasoning intensity controls ranging from low to high
- ▪Google surfaced a new Ultra mode within Google AI Studio promising advanced tools, while Logan Kilpatrick confirmed efforts to optimize Argon model performance
Recursive self-improvement in AI
- ▪OpenAI and Anthropic have reported similar progress in using artificial intelligence to build new models
- ▪Google DeepMind employee Vedant Misra attributed the accelerated iteration speed of Google's new Gemini 4 models to recursive self-improvement using artificial intelligence
Debatable claims
- ▪Google's Gemini 4 models match Anthropic's flagship capabilities in agentic programming
- ▪Recursive self-improvement is the primary driver of rapid AI model iteration
Story comments
Loading comments…