Business Insider: before Gemini 4 Argon fully ships, Google staff are testing a stronger internal build codenamed Carbon that one says codes like Opus 5.5

Why it matters

【Key points】

Business Insider reported on Oct 9 (US time) that Google employees are testing a new Gemini 4 version codenamed Carbon. The report is based on internal documents, screenshots and staff messages seen by the reporter: Carbon appeared in recent days on Jetski, Google's internal coding platform, and may be a new checkpoint (a saved mid-training version) of Argon or a separate Gemini 4 model. One employee said it "feels like Opus 5.5" for coding, while cautioning it needs more testing. Google declined to comment.

【Verifiable】

The story is by Hugh Langley, who covers Google. Details worth noting: an internal document shows Gemini 4 models tested under the codenames Argon, Barium and Carbon, and internal names don't map to public ones, for example the model chosen to ship as Argon was internally "Barium-B"; one employee compared early Argon builds on some coding tasks to the older Claude Opus 5; Google is not guaranteed to release these updates. Argon was announced on Sept 30 but so far goes first to partners in Google's Fairwind cybersecurity testing program, not a full public rollout. Related posts on X totaled about 280 over the past 24 hours, running above 50 an hour in the two hours after the story and 10 to 20-plus an hour since; the most-shared repost has about 37,000 views.

【Our read】

What the report really shows is not how strong Carbon is but that Google's bottleneck has shifted from building models to releasing them: ten days after announcing Argon it is still in limited testing, while a stronger build is already running internally. "Feels like Opus 5.5" is a staff impression and is not a capability result until public evals or developer access exist. The more useful point for developers is that codenames and public names are decoupled: Carbon could well ship as an update under the Argon name, same label, different capability. Third-party evals and enterprise buyers will need to track checkpoints and dates, not just model names.