An internal assessment reportedly found that Carbon, a version of Gemini 4 said to be in development at Google, has programming capabilities close to 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 Real-World Coding Work
The focus of this report is not an official announcement of a new model, but internal testing at Google.Reference articlereports that several Gemini 4 development builds, including Argon, Barium, and Carbon, are being tested, based on documents, screenshots, and internal communications reviewed by Business Insider.
Carbon has been integrated into Google’s internal coding platform, “Jetski,” and is said to outperform Argon mainly on programming tasks. One employee described the experience as similar to using Anthropic’s Opus 5.5. By contrast, another employee said the early version of Argon reminded them of Opus 5 on some programming tasks.
It is important to understand what “comparable” means here. The report cites an employee’s experience using the model, not the results of a test in which both models solved the same problems under the same conditions. It provides no figures, such as accuracy rates or task completion times, so it does not establish that Carbon and Opus 5.5 perform equally. The employee who evaluated it also said more testing is needed.
Development Codenames and Product Differences Are Not the Same Thing
The names Argon, Barium, and Carbon might suggest three products with different purposes. However, the reference materials indicate that Barium and Carbon may be works in progress or updated versions in the Argon lineage. It is not clear whether Carbon will become a standalone product or be released as an update to Argon.
The evolution of the codenames supports that interpretation. The materials say Argon was previously known as Barium-B, and note that Carbon also had an internal label suggesting it was the model corresponding to the next Gemini Pro. A change in development codename should be distinguished from the addition of a new product category for users.
The materials also describe Google’s stated roles for 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 not as a model optimized for speed, but as one designed to handle difficult tasks.
Separating AI-Assisted Development From Self-Improvement
As a possible explanation for the rapid updates to development builds, the reference article highlights efficiency gains from using AI in model development. Google DeepMind employee Vedant Misra also reportedly referred to recursive self-improvement in a post on X reacting to the report.
In this context, self-improvement means using AI to develop more capable AI. However, the post alone does not reveal where AI was used in Carbon’s development process or how much time it may have saved. Nor does it say that Carbon improved itself without human involvement.
The pace of model updates and the technical methods used to accelerate them are separate pieces of information. The materials report that Google is testing multiple development builds and evaluating them in its internal work environment. The reference to self-improvement offers a clue to the broader context, but is not presented as an experimental result identifying the cause of performance gains.
App Changes Are Underway, but Availability Details Remain Unannounced
Changes to user-facing interfaces have also been reported. The reference materials say the Gemini app displays an “Automatic” mode for the 3 series, along with adjustable reasoning strength from low to high, while Google AI Studio has added an “Ultra” mode touting advanced features and tools.
These changes are presented as preparations for the launch of Gemini 4, but mode names displayed in an interface alone do not confirm that Carbon is available. Google’s Logan Kilpatrick reportedly said he was working to bring out the full capabilities of Argon.
For developers, the key question is how the positive internal assessment will translate into a model they can actually use. The reference materials do not include Gemini 4’s official release date, Carbon’s final product name, pricing, access terms, or comparable performance metrics.
Comments