Early tests of Google Gemini 4 Carbon show 3D improvements

Early access for Carbon

Google appears to be moving quickly to release its next major AI checkpoint, Gemini 4 Carbon. While the Gemini 4 Argon model has only been public for about a week, the company is already preparing the successor for a wider audience. This move follows reports of Carbon being tested inside Google's internal Jetski coding tool earlier this month.

According to reports from TestingCatalog AI News, a specialized early access program has begun for Carbon. This program is designed for a specific group of developers, researchers, and partners who receive access to upcoming models before the general public. These users are governed by a Trusted Tester Agreement that requires them to keep benchmarks and outputs private.

Despite these privacy rules, several outputs from the model have already surfaced on social media. Users have shared results from the LM Arena, a platform where models are compared against one another. These early glimpses provide a preview of how Carbon performs compared to the current standard set by Argon.

Performance and 3D capabilities

Initial feedback from early testers highlights a significant shift in how the model handles complex spatial tasks. Specifically, Carbon is being described as noticeably better than Argon when performing 3D work. One specific example of this is the voxel pagoda comparison, where Carbon demonstrated higher quality than its predecessor.

The model also underwent the classic SVG pelican on a bicycle test to gauge its creative capabilities. While the results were positive in some areas, testers noted that the model still struggles with certain types of visual tasks. Specifically, the output for frontend and UI design was described as still being weak by those who tested it.

Another notable characteristic of Carbon is its behavior during complex prompts. Testers reported that the model feels "less lazy" than Argon. This suggests that while it may not be perfect, it is attempting to provide more complete and thorough responses to difficult instructions.

The trade-off of speed

The improvements in quality and 3D handling come with a significant cost in processing time. Early testers have described Carbon as being "incredibly slow" during various evaluations. Depending on the specific task being performed, the model can be roughly two to four times slower than the Argon version.

This discrepancy in speed is a critical factor for developers who need to balance output quality with latency. While a slower model might provide a more accurate or sophisticated result, it may not be suitable for applications requiring near-instantaneous responses. Google will need to find a balance as they move toward a public release.

The urgency for these updates may be driven by the massive amount of data being processed by Google's current systems. Following the launch of Argon, Google processed around 25 trillion tokens in just the first 72 hours. This high volume of activity suggests a strong demand for high-performing models that can keep up with user needs.

The broader AI landscape

Google is not operating in a vacuum as it prepares the release of Carbon. The AI industry is currently seeing a rapid cadence of new model checkpoints from several major players. For instance, OpenAI is reportedly preparing GPT-6.1 Terra, which has already surfaced in Microsoft code, while Anthropic is reaching partners with Fable 5.5.

Google's strategy seems to involve building a steady pipeline of improvements to maintain its competitive edge. Beyond the chatbot interface, the company is also exploring broader permissions for computer use within its desktop app. These developments suggest that Gemini 4 is the foundation for a wider suite of integrated tools.

The company is also reportedly working on Nano Banana 3 Pro, which is expected to be built on top of the Gemini 4 architecture. With a potential window of November to December in view, Google is positioning itself to offer a variety of model sizes and capabilities. The success of Carbon will depend on whether it can fix Argon's weaknesses without making users wait too long for results.