Google is preparing to publicly roll out Gemini 4 Argon while employees at the company have begun testing a newer, possibly more powerful version over the past few days. This version could further narrow Google's gap with its rivals in the AI race.
Based on documents and images seen by Business Insider, this version of Gemini 4 is being tested under the internal codename “Carbon.” “Carbon” is likely a fresh update to the Argon model, also known as a “Checkpoint.” Over the past few days, this model has been made available to employees on Jetski, Google’s internal coding platform.
One employee told Business Insider that “Carbon” delivers a coding experience “similar to Opus 5.5.” Opus 5.5 is Anthropic’s most powerful model and is designed for coding tasks that require agentic coding capabilities and long-running operation execution. However, the same employee emphasized that confirming this impression requires more testing.
These tests indicate Google is rapidly developing the next generation of Gemini 4 models. Google introduced Gemini 4 Argon on September 30. This model is larger than Google’s prior advanced “Pro” models, and the company has described it as comparable to leading competitors’ models in coding and cybersecurity, including Anthropic’s Opus and OpenAI’s Astra. Google also said in the announcement that Argon lags competitors on some metrics and did not outperform them on two of four programming-related tests.
Argon is not yet available to the public. Google has only provided it to a number of cybersecurity partners and is participating in the U.S. government’s voluntary pre-release access process for models. The company has not announced a public release schedule. Google set Argon’s introductory price at $2 per 1 million input tokens and $10 per 1 million output tokens. After the introductory period ends, those rates will rise to $4 and $20. Sundar Pichai, Google’s CEO, also set the model’s output cap at 1 million tokens and called it “industry-leading.” He said internal Google teams are using Argon in areas from coding to quantum computing.
The testing of “Carbon” shows Google’s intense competition to build AI agents specialized in programming. Anthropic and OpenAI, with models designed for complex engineering tasks, have so far paved the way for developers, and Google is trying to close that gap. The company also said it no longer plans to release Gemini 3.5 Pro, a model Pichai had initially promised in June.
The initial report was published by Business Insider. Google has not yet commented officially on the “Carbon” model, and its details could change before release.