Google's race to catch up in AI-assisted coding may have a new contender — and it is not the model everyone is waiting for. Business Insider reported this week that Google employees are internally testing a Gemini 4 variant codenamed Carbon, and early feedback suggests it could be the strongest coding model the company has built.
Carbon was made available in recent days on Jetski, Google's internal coding platform. One employee who tested it told Business Insider its coding performance "feels like Opus 5.5" — Anthropic's flagship model for long-running coding tasks. Another wrote in an internal channel simply, "Carbon is really good!" The employee who made the Opus 5.5 comparison cautioned that further testing would be needed to assess its capabilities across a broader range of tasks. Google declined to comment on the report.
Documents reviewed by Business Insider indicate that Google has been evaluating several Gemini 4 checkpoints under the codenames Argon, Barium, and Carbon. One document identified Barium-B as the checkpoint selected for release under the Argon name — the model Google officially announced on September 30 and began rolling out to selected cybersecurity partners through its Fairwind Program. Whether Carbon becomes an Argon update or a separate Gemini 4 model remains unclear, and Google may never release it publicly.
The testing underscores how central coding has become to the model wars. Earlier versions of Argon drew positive feedback from employees, but some testers felt the model lagged rivals on certain coding tasks — one compared an earlier build to Anthropic's older Opus 5 on some jobs. Google pushed back on that framing, telling Bloomberg it would be "inaccurate to say that Gemini 4 is underperforming in areas such as coding," with one employee telling the outlet there is "large consensus" internally that Gemini 4 is at the frontier.
Public signs of progress are appearing too. Antigravity, Google's development tool, now references Gemini 4 Argon in its model selector with three context configurations, and the Gemini web app recently added low, medium, and high reasoning-effort settings — likely groundwork for Gemini 4's arrival. With OpenAI and Anthropic both pressing hard on agentic coding, Carbon looks like Google's insurance policy: a sharper checkpoint waiting in the wings while Argon takes the spotlight.