Google staff test new AI codenamed Carbon that "feels like Opus 5.5"

Last month, Google had to swallow its pride and let its engineers use Anthropic's Claude to write code. But reinforcement from within Mountain View might already be taking shape.

According to an internal report from Business Insider, Google staff have begun testing a new Gemini 4 variant internally codenamed Carbon. The model was rolled out to Jetski, Google's proprietary internal coding platform, over the past few days.

Early internal reactions are surprisingly glowing. In internal messaging channels, one employee reported that Carbon "feels like Opus 5.5" when handling coding tasks. Another simply posted, "Carbon is really good!"

The confusing alphabet soup of codenames

Google has a long history of decoupling its internal project codenames from what it eventually sells to the public, and Gemini 4 is no exception.

Back on September 30, Google officially announced Gemini 4 Argon. However, internal documents reveal that the public Argon model was actually built on an internal checkpoint called Barium-B.

Carbon represents a newer training checkpoint developed right after Barium. Internal threads show employees discussing whether Carbon will be rolled out as an incremental patch under the Argon label, or packaged as an entirely separate entry in the Gemini 4 lineup. Google declined to comment on its internal roadmap.

Playing catch-up on code

It is no accident that staff immediately compared Carbon to Anthropic's flagship. Software engineering has long been Gemini's most glaring vulnerability.

Bloomberg previously reported internal grumbling among Google staff, with engineers complaining that Argon put up flattering benchmark numbers while stumbling on real-world coding duties. One staffer even likened early Argon iterations to the much older Claude Opus 5.

Independent benchmarks tell the same story. On Terminal Bench 4, Argon posted a score of 57%, trailing behind OpenAI's GPT-6 Astra at 59%, Claude Opus 5.5 at 60%, and Claude Sonnet 5.5 leading the pack at 64%.

That performance gap forced Google to lift internal restrictions in September, officially letting company engineers lean on Claude to get their day jobs done.

Meanwhile, everyday developers still cannot access Argon anyway. Google has only distributed it to select cybersecurity partners under its Fairwind Program to test cyber vulnerabilities, leaving paying API users and Google AI Ultra subscribers waiting in line.

Why it matters

For developers, Carbon is a promising signal that Google might finally fix its coding woes. If Carbon truly matches Opus 5.5, developers relying on Google Cloud and Gemini tools will get an overdue boost. Still, we think a heavy dose of skepticism is warranted: "feels like" is an anonymous vibe check, not a verified benchmark, and Google hasn't committed to any public release date.

For the wider AI race, coding agents have become the ultimate testing ground for frontier reasoning. Anthropic and OpenAI have captured immense goodwill among technical users, and Google's decision to let its own workforce rely on Claude was a glaring concession of defeat.

Testing Carbon on Jetski shows Google is moving fast to stop the bleeding. But in software engineering, developers care about reliability, not chemistry-inspired codenames. Google will actually need to ship this thing if it wants credibility back.