An internal assessment reportedly found that Carbon, a version of Gemini 4 said to be in development at Google, has programming capabilities close to those of Anthropic’s Opus 5.5. It has been tested in Google’s internal development environment and is said to outperform another development build, Argon, mainly on programming tasks. The reference materials do not provide an official release timeline or figures from comparative tests.
Carbon’s evaluation comes from hands-on coding work
The focus of this report is not an official announcement of a new model, but its testing within Google.The cited articlereports that Google has been testing several Gemini 4 development builds, including Argon, Barium, and Carbon, based on documents, screenshots, and internal communications reviewed by Business Insider.
Carbon has been deployed on Jetski, Google’s internal coding platform, and is said to outperform Argon mainly on programming tasks. One employee described using it as similar to Anthropic’s Opus 5.5. Another employee said that an early version of Argon reminded them of Opus 5 for some programming tasks.
The key point is what “comparable” means here. The reports describe an employee’s experience using the model, not the results of a test in which both models were given the same problems under the same conditions. No figures for metrics such as accuracy or task completion time are provided, so it has not been established that Carbon performs as well as Opus 5.5. The employee who evaluated it also said further testing is needed.
Development codenames and product distinctions are separate issues
The names Argon, Barium, and Carbon might suggest three products with different purposes. But the reference materials indicate that Barium and Carbon may be in-development versions or updates in the Argon family. It remains unclear whether Carbon will become a standalone product or be released as an update to Argon.
The evolution of the codenames supports that view. According to the materials, Argon was previously called Barium-B, and Carbon also had an internal designation suggesting it was the model corresponding to the next Gemini Pro. A change in a development codename should be distinguished from the addition of a product category for users.
The materials also describe Google’s positioning of its models: Argon for advanced reasoning, Flash for speed and high-volume processing, Omni for media generation, and Gemma as a lightweight model that can also run on devices. Reasoning refers to organizing multiple conditions and working through them step by step to reach an answer. Argon is described as a model for tackling difficult tasks, rather than one that prioritizes speed.
Separating AI-assisted development from self-improvement
As a possible reason for the rapid updates to development builds, the cited article points to greater efficiency in model development using AI. Google DeepMind employee Vedant Misra also reportedly referred to recursive self-improvement in a post on X responding to the report.
In this context, self-improvement refers to using AI to develop more capable AI. The post alone, however, does not show where AI was used in Carbon’s development process or how much time it saved. Nor does it say that Carbon improved itself without human involvement.
The pace of model updates and the technical mechanisms that accelerated them are separate matters. The materials describe efforts to test multiple development builds and evaluate them in Google’s internal working environment. References to self-improvement offer a way to consider the background, but are not presented as experimental results identifying the cause of performance gains.
App changes are underway, but release details remain undisclosed
Changes to user-facing interfaces have also been reported. The reference materials say the Gemini app displayed an “Automatic” mode for the Gemini 3 series and a reasoning-strength setting adjustable from low to high, while Google AI Studio showed an “Ultra” mode touting advanced features and tools.
These changes are presented as preparations for the launch of Gemini 4, but the names of modes shown in an interface do not confirm that Carbon will be released. Google’s Logan Kilpatrick reportedly said he was working to bring out Argon’s full capabilities.
For developers, the question is how the positive internal assessments will translate into a model they can actually use. The reference materials do not give an official release date for Gemini 4, Carbon’s final product name, pricing, terms of use, or performance metrics that would allow for comparison.
Comments