Google tests internal Gemini 4 model Carbon for coding

Google employees are testing an internal checkpoint model codenamed Carbon for the Gemini 4 family, reportedly with improved coding performance, according to Business Insider. The tests take place on the company's internal coding platform Jetski.

Internal documents indicate Google has tested multiple element-named codenames during Gemini 4 development, including Argon, Barium and Carbon. The internal name Barium-B was chosen as the final checkpoint for the public Argon model, while Carbon is specified separately as a model with significantly raised performance.

Employees who have tried Carbon say it reaches a level comparable to Anthropic's top model, with one describing its coding ability as similar to Opus 5.5. A recent update to Google's AI coding environment Antigravity showed model selector context options of 256K, 512K and 900K, with thinking effort levels of low, medium and high.

It is not yet confirmed whether Carbon will be delivered as an update to Argon or released as a separate standalone model in the Gemini 4 family.

Argon's public rollout

Google officially announced Argon on September 30 and gave select cybersecurity partners early access. The company said Argon was integrated into its enterprise Gemini Agent announced on the 8th, with deployment to some partners beginning. Google is expected to begin the full launch of Argon to API customers and AI Ultra subscribers in October.