Google’s Gemini 4 Carbon Reportedly in External Testing
- •Gemini 4 Carbon is reportedly being tested with external partners on the LM Arena model-comparison platform.
- •Early users say its 3D work improves on Argon, while task times are 2–4 times longer.
- •Its performance and release schedule have not been officially confirmed.
Google’s Gemini 4 Carbon, a model in development, was reportedly being tested with external partners on the AI model-comparison platform LM Arena, according to a report on the 23rd local time. TestingCatalog cited posts by test participants. Early users said Carbon produced better 3D results than Gemini 4 Argon, but responded more slowly and fell short at webpage design.
Earlier this month, Carbon was reported to be under internal testing by employees using Google’s Jetski coding tool. Business Insider reported its existence based on internal documents and screens. One employee said its coding ability “feels like Opus 5.5,” while adding that more testing was needed. It remains unclear whether Carbon is an update to Argon or a separate model.
The leak-focused X account Lentils claimed on the 21st that Carbon’s Early Access Program (a limited pre-release testing program) had begun and that release was near. On the 22nd, it shared results from an updated test version. Lentils said Carbon was noticeably better at 3D tasks than Argon but was not perfect; it responded more slowly while seeming more willing to engage with tasks. Another leak account, Lyra, shared a screenshot of an Early Access Program document saying invited developers, researchers, and partners could use Gemini API and AI Studio models before release.
The X account Lumina published a comparison of “Voxel Pagoda” results, in which small cubes are stacked to make a tower, against other advanced models. Lumina said Carbon’s output quality was clearly higher than Argon’s, but each task took 2–4 times as long. It also said Carbon fell short at webpage appearance and user interface (UI) generation, while leaving a generally positive impression across several tested tasks.
The Early Access Program document alone does not officially confirm Carbon’s external testing. The published results are early assessments of tasks selected by users, and the version made available for testing is also uncertain. Carbon’s performance and release schedule have not been officially confirmed.