Google's Gemini 4 Carbon is already being tested behind the scenes

What's new

Google appears to have a new model on the way: Gemini 4 Carbon. There is no official announcement yet, so everything we know comes from leaks and posts by testers. According to the user Lentils on X, an early access program has started in which partners have been allowed to try the model. It is also being tested on LMArena, where users compare AI models in anonymous head-to-head matchups.

Lentils expects a launch to be very close. That is only an expectation, not a confirmed date. Carbon's final public name has not been confirmed either. Business Insider reported around October 9, 2026, that Google employees are also testing the model internally in the tool Jetski.

What's smart

The strength testers highlight most is 3D. Lentils judges Carbon to be noticeably better than Argon (the previous model, which Google announced on September 30, 2026), but not perfect. The tester Lumina compared a pagoda built from 3D blocks (voxels) against other leading models and believes the quality is higher than Argon's.

For coding, there is only one impression: According to Business Insider, an anonymous Google developer said the model's coding ability felt comparable to Anthropic's Opus 5.5. There is no comparison with Argon on code, and no other source has confirmed the statement.

Lentils also writes that the model is less lazy than Argon, meaning more willing to finish the task.

Cheaper or better?

It appears to be better in some areas, but whether it is cheaper is unknown. Carbon's price has not been disclosed, and there are no public benchmarks (standardized tests).

There are also downsides. Lumina calls Carbon "incredibly slow," roughly two to four times slower than Argon depending on the task. Lentils says the frontend and user interface are still poor. Lumina also found the interface weak in a classic test where the model has to draw a pelican on a bicycle, but otherwise thought it handled many tasks well.

For comparison, Argon costs $2 per million input tokens and $10 per million output tokens during an introductory period (tokens are chunks of text). Afterward it will be $4 for input and $20 for output. According to the blog eesel.ai, that matches Claude Opus 5.5. The same blog calculates that, despite having the same introductory price as GPT-6.1 Sol, Argon costs 2.7 times more per task because of its heavy token usage. That is the blog's own calculation (via Artificial Analysis).

What it's good for

According to the testers, Carbon is especially interesting for 3D tasks. For general use, it is too early to say anything, and the speed may be a drawback if you need fast answers.

Take the statements with a grain of salt. TestingCatalog points out that the posts are based on arena matchups and hand-picked tasks, not benchmarks, and that nobody knows which version the testers are being served. According to TestingCatalog, Google's agreement for early testers also requires that unreleased models be kept private, so posts may be removed or changed.

The sources do not say whether or when Carbon will become available in Denmark and the EU, or what it will cost regular users.