Google Launches Gemini 4 Argon

Google has introduced Gemini 4 Argon, an AI model with a token limit of 1 million, designed for deep problem solving and complex workflows. The model is currently being rolled out to trusted cyber defenders through the Fairwind Program, with Wiz among the first companies to utilize it for cybersecurity defense.

Gemini 4 Argon is priced at $2 per million input tokens and $10 per million output tokens, with a significant discount for cached input tokens at 95% off. The model has freed up over 300 TiB of memory, with estimated total savings ranging from 500 TiB to 1 PiB. It also supports a codebase migration size of over 800K lines for the Fuchsia Zircon kernel.

Performance metrics for Gemini 4 Argon include a score of 77.9% on DeepSWE v1.1 and leading results across finance, coding, legal, and tax work on the Vals Index. The model achieved a score of 51.3% on AutomationBench and 91.7% on LVBench, with a notable improvement on CWE-bench v1, scoring 68% compared to its predecessor's 3.8%. Additionally, it demonstrates a 40% performance improvement in quantum algorithmic optimization and is 2.7x faster than the Rust port for libgav1.

Google claims that Argon delivers frontier performance in complex workflows and is actively engaged in the U.S. government’s voluntary process for pre-release model access. To mitigate risks, the model is designed to refuse harmful requests to prevent misuse in cyber or CBRN attacks and is noted as the most resilient model against indirect prompt injections to date. A gradual public release is planned following feedback from early testers.