Google Prepares Gemini 4 'Argon' Release as Employees Test Upgraded 'Carbon' Model
Google is reportedly preparing to launch its Gemini 4 'Argon' AI model while internally testing an upgraded checkpoint codenamed 'Carbon' on its internal coding platform, Jetski. Leaked internal feedback indicates that Carbon demonstrates advanced programming capabilities that rival Anthropic's flagship Claude Opus models. This development highlights Google's rapid iteration pace in AI development as it seeks to challenge Anthropic's dominance in complex, long-horizon software engineering tasks. If Carbon's capabilities hold up in rigorous benchmarks, it could significantly advance automated code generation and agentic developer tools. Internal documents show Google has tested several Gemini 4 variants, including Argon, Barium, and Carbon, with early Argon builds performing on par with Claude Opus 5 and Carbon reaching performance compared to Opus 5.5. Because internal codenames differ from public branding—such as Argon being designated as 'Barium-B' internally—it remains uncertain whether Carbon will release as a fine-tuned Argon update or a standalone model.
## BACKGROUND
Jetski is Google's proprietary internal software development platform where company engineers test unreleased AI models on real-world coding workflows. Software engineering benchmarks like DeepSWE evaluate AI agents on long-horizon, complex tasks within isolated environments to measure autonomous problem-solving capabilities.