Google unveils Gemini 4 Argon AI model, pending external access

Google said its new Gemini 4 Argon model delivers what it calls industry-leading performance in software development, knowledge-intensive tasks and cybersecurity, though the model is not yet accessible to external users. The company indicated that internal engineers are already employing Argon, noting that it helped reduce data-center memory usage by about 300 terabytes and assisted in converting large C/C++ codebases to Rust, including significant portions of the re2 and libgav1 libraries and over 800,000 lines in the Fuchsia OS Zircon kernel.

Google cited benchmark results to support its claims, saying Argon achieved a 77.9 percent score on the DeepSWE v1.1 software-engineering test, surpassing competing models such as GPT-6 Astra, Fable 5.1 and Opus 5.5. It also highlighted a leading result on the Vals Index economic-analysis benchmark. The model is said to support up to one million tokens per request, a substantial increase from the 64,000-token limit of earlier Gemini versions.

The initial rollout focuses on cybersecurity, with a limited group of trusted testers in the company's Fairwind Program receiving early access. One partner, Wiz, reportedly used Argon to identify a critical vulnerability affecting hospital systems worldwide, a flaw that Google claims other frontier models missed. Google said Argon incorporates monitoring systems that track the model’s reasoning process and can intervene if outputs deviate from expected bounds, emphasizing the importance of reasoning transparency.

Google has not set a public release date or disclosed pricing for an API. It indicated that after the cybersecurity testing phase, broader availability will be offered to paid API customers and subscribers to its Google AI Ultra service.