Google introduced the Gemini 4 Argon model, but the service is not yet open to external developers. The company says internal engineers have already begun extensive use of the system. According to Google, Argon leveraged fleet-wide telemetry data to reduce memory consumption in its data centers by roughly 300 TiB. The announcement follows the earlier Gemini 3.5 Pro release, which remains superseded by the new model.
Argon’s internal rollout includes a large-scale effort to translate Google’s C and C++ codebases into Rust. Engineers have targeted key libraries such as re2 and libgav1, moving thousands of lines of code, while the Fuchsia OS Zircon kernel has seen more than 800,000 lines converted. Google presents the migration as a step toward higher safety and performance across its software stack.
The company released a series of benchmark results to illustrate Argon’s capabilities. On the DeepSWE v1.1 software-engineering test, Gemini 4 Argon achieved a 77.9 percent score, surpassing competitors identified as GPT-6 Astra, Fable 5.1 and Opus 5.5. Google also highlighted an industry-leading result on the Vals Index economic-analysis test, positioning Argon as a leading performer on long-horizon tasks.
Despite remaining in limited testing, Google disclosed pricing for an upcoming API. The model will cost $2 per million input tokens and $10 per million output tokens, with a 95 percent discount on cached input tokens. Argon also raises the maximum output length to one million tokens, a substantial increase from the 64,000-token ceiling of earlier Gemini versions, enabling more complex single-step operations.
The initial deployment of Argon concentrates on cybersecurity applications. Google plans a phased rollout to a small cohort of trusted participants, including partners in its Fairwind Program. Security firm Wiz is reported to have employed Argon to identify a critical vulnerability in a hospital-system platform that could expose personal data worldwide, a flaw that other frontier models allegedly missed.
Google emphasized safety mechanisms built into Argon to address concerns about model misalignment. The system monitors the model’s chain-of-thought processes and can interrupt execution if the reasoning deviates from predefined bounds. According to the company, this approach underscores the importance of reasoning transparency in the development of frontier AI models.