Google Tests Gemini 4: Carbon Compared to Opus 5.5

Google is evaluating several variations of Gemini 4 internally, with no release date in sight. Carbon is reportedly being tested on the in-house coding platform and is approaching the performance level of Anthropic's Opus 5.5 in programming. Meanwhile, changes are already appearing in Google's consumer and developer tools.

No date, but products already adjusted

No official launch date has been announced for Gemini 4. Nevertheless, Google is beginning to adapt its applications. The Gemini app now features an "Automatic" mode for the series 3 models, with adjustable reasoning intensity ranging from low to high. Google AI Studio has introduced an "Ultra" mode that promises "advanced skills and tools." Logan Kilpatrick from Google confirmed that the team is working to make the most of Argon. Vedant Misra, an employee at Google DeepMind, mentioned on X the concept of "self-recursive improvement," referring to the use of AI to develop better AI. OpenAI and Anthropic have also reported similar advancements in this area.

Carbon tested internally, compared to Opus 5.5

According to internal documents and discussions reviewed, Google is testing several variants of Gemini 4, including Argon, Barium, and Carbon. Carbon was recently deployed on Jetski, Google's internal coding platform, and is said to outperform Argon in programming tasks. A Google employee compared Carbon to Anthropic's Opus 5.5 coding model, noting that further testing is needed. Thus, Carbon would match the performance of Opus 5.5 for coding. In contrast, early versions of Argon reminded another employee of the older Opus 5 in certain tasks, although internal feedback on Argon has been generally positive.

Clarified roles and evolving identity of Argon, Barium, and Carbon

Argon has not yet been launched. Initial information linked it to the Flash range, which specializes in speed, but this hypothesis was abandoned after the Gemini agent was unveiled for Google Workspace. Google is defining specific functions for each of its models: Argon focuses on advanced reasoning, Flash on speed and handling large volumes, Omni on media generation, and Gemma on lightweight tasks with open weights. According to Google, Argon represents its most sophisticated reasoning model, akin to Opus at Anthropic or Astra at OpenAI. Carbon and Barium are likely intermediate versions or evolutions within the Argon series, rather than independent variants. A staff member described Carbon as the "next Gemini model," which aligns with internal files indicating that the previously presented Argon model was named Barium-B. It has yet to be determined whether Carbon will be released as an update to Argon or as a distinct model. The rapid pace of updates, similar to that observed with the Flash models, suggests that Google has improved its use of AI to design new models.