Google unveils Gemini 4 Argon AI with industry-leading code and security chops—but no public access yet

Google has introduced Gemini 4 Argon, a new AI model designed to tackle complex challenges in software engineering, knowledge work, and cybersecurity—but the general public won't be able to try it yet. The company is currently running limited tests with the model internally, while public access remains unavailable.

Performance Claims#

Gemini 4 Argon is built on benchmark results that Google says demonstrate its competitive edge. On the software engineering DeepSWE v1.1 test, the model achieved 77.9% performance, outpacing comparable offerings like GPT-6 Astra, Fable 5.1, and Opus 5.5. Google also highlights strong performance on long-horizon tasks and economic analysis benchmarks, positioning Argon as a frontier-class model.

Real-World Deployment Inside Google#

Google's own teams have already begun leveraging Argon for internal work. The model has helped the company optimize data center memory usage, recovering 300 TiB across its infrastructure using fleet-wide telemetry analysis. Argon-powered agents are also handling code migration projects, including translating C/C++ libraries to Rust—work that spans thousands of lines in critical systems like the re2 and libgav1 libraries, as well as over 800,000 lines in the Fuchsia OS Zircon kernel.

Capabilities and Specifications#

One standout feature is Argon's context window: the model supports up to 1 million output tokens, a significant jump from the 64,000-token limit in earlier Gemini versions. This expanded capacity enables users to request more comprehensive outputs and handle larger tasks within a single interaction. Pricing for API access has not yet been announced, since the model remains in limited testing phases.