Rumor: Google's "Gemini 4 Carbon" in testing, better at 3D but slower

What this is about

Google appears to be working on a new model in the Gemini 4 family, codenamed "Carbon." According to the outlet Testingcatalog, citing posts from testers, the model has been rolled out to some external partners since October 23, 2026 (local time) and is also being evaluated on LM Arena, a platform where users compare AI models against each other anonymously. This comes about a week after "Argon," another Gemini 4 model, appeared. Google hasn't confirmed anything officially.

What we knew so far

In early October, Business Insider reported on Carbon, citing internal documents and screenshots. Google employees were testing the model at the time in the internal coding tool "Jetski." One employee said the coding ability felt "like Opus 5.5" (a model from Anthropic), but stressed that more testing was needed. It's still unclear whether Carbon is an update to Argon or a standalone model.

What early testers report

The leaker account "Lentils" on X wrote on October 21 that Carbon's Early Access Program (EAP, an advance access scheme for invited users) had started and that the release was near. On October 22, it showed results from an updated test version. Carbon is noticeably better than Argon at 3D tasks, but not perfect. Google apparently is focusing on this area first, before frontend and UI follow. The impression: the model responds more slowly but tackles tasks with more determination.

The account "Lyra" shared a screenshot of the documentation for the Gemini API Early Access Program. According to it, invited developers, researchers and partners can use models in the Gemini API and in AI Studio before release. That doesn't prove, however, that Carbon is being tested externally.

The account "Lumina" compared a "voxel pagoda" (a pagoda built from small cubes) with results from other frontier models. The quality was significantly higher than with Argon.

Weaknesses: speed and web design

Testers agree on speed: according to Lumina, processing took 2 to 4 times as long as with Argon, depending on the task. There are still shortcomings in website visuals and user interfaces (UI). Overall, though, Lumina's impression across the many tasks tested was good.

Our take

This report is secondhand rumor: the results come from individual users' self-chosen tasks, and it's not certain which test version was actually rolled out. Performance and release date are not officially confirmed. It's still interesting, because it shows that Google is testing several Gemini 4 variants in parallel and apparently improving 3D and coding abilities first, at the expense of speed. Whether the model holds up in everyday use can only be judged after release.