Google tests new Gemini 4 Carbon checkpoint with improved 3D and coding

A new checkpoint for Carbon

Google recently unveiled Gemini 4 Carbon, a model currently restricted to a small subset of cyber defenders. Despite the limited access, the model has already begun appearing in public testing environments like Arena. These early tests suggest that Google is actively refining the model's performance before a broader release to the general public.

Testers on the Arena platform have identified a new checkpoint that appears to be a significant upgrade over previous versions. One user noted that the new build feels noticeably faster and more capable than what was available just weeks ago. This rapid iteration suggests that Google is moving quickly to address early feedback regarding the model's speed and output quality.

The new build, which some testers refer to as Carbon-H, is reportedly being served under an Argon label. Early reactions to this specific checkpoint have been more positive than the initial reception of the Carbon series. Users are beginning to notice more sophisticated outputs in specific categories like 3D design and complex coding tasks.

Advancements in 3D and logic

One of the most prominent improvements in the new Carbon checkpoint is its ability to handle 3D rendering tasks. In side-by-side comparisons, Carbon produced cleaner models using rounded shapes and diamond pieces, whereas the Argon model assembled objects from simple blocks. This indicates a higher level of spatial awareness and aesthetic refinement in the Carbon architecture.

The model also demonstrates a high level of reasoning, sometimes taking between 8 and 10 minutes to think before it generates a response. This deliberate processing allows it to handle complex prompts that require deep planning rather than immediate generation. Some testers have even compared the quality of these outputs to those of other high-end models like Opus 5.5.

In practical applications, this reasoning power is already being utilized by security companies. The head of Wiz reported that they have used the model to work out complex exploit chains. He noted that they have never had so many novel discoveries while bug hunting, highlighting the model's utility in high-stakes technical environments.

Remaining hurdles for developers

Despite the progress in 3D and logic, the model still faces challenges in specific areas of software development. Testers have noted that front-end coding remains a sore spot for the current iteration. Users are hoping that Google will eventually fix issues related to web development and 'slop' user interfaces in future updates.

While the 3D work is considered a major strength, the model was previously criticized for being painfully slow and weak at front-end tasks. The latest tests show that the speed issues are being addressed, but the core competency in web development has not yet reached the same level as its 3D capabilities. This creates a mixed experience for developers who need a versatile all-in-one tool.

In animation tests, the model received high marks for movement in a pixel-style animation task. However, because these ratings come from individual prompts, they serve as a rough guide rather than a definitive benchmark. The community is watching closely to see if Google can balance these strengths to create a more well-rounded experience for all types of creators.

Pricing and future integration

Google has already established a pricing structure for the upcoming wider release of Gemini 4 Carbon. It will be offered at the same introductory price as the Argon model, which is $2 per million input tokens and $10 per million output tokens. This pricing strategy aims to make the high-performance model accessible to a larger audience of developers and consumers.

The model is also being positioned as a potential backbone for other Google projects currently in development. For instance, Google is preparing for the cloud execution of Antigravity as the Jules project heads toward a shutdown on December 31. A faster and stronger version of Carbon could be a natural fit for these infrastructure needs.

Additionally, DeepMind recently previewed Genie 4, which features native audio capabilities. As Google continues to refine Carbon, the company is looking to integrate these improvements across its ecosystem. If Google can successfully close the gap in front-end development, the wider release of Carbon is expected to be well-received by the tech community.