Google Unveils Frontier Model 'Gemini 4 Argon'... Supports 1 Million Output Tokens

Google has unveiled its new frontier model 'Gemini 4 Argon'. This model delivers frontier-level performance across various workflows, including complex software engineering, legal and financial tasks, and cybersecurity defense. It is currently being deployed first to trusted cybersecurity experts, with access planned to be gradually expanded to developers and enterprises in the future.

Performance and Pricing

Argon is offered at a launch special price of $2 per million input tokens and $10 per million output tokens. Cached input tokens receive a 95% discount. Additionally, the model's output token limit has been significantly expanded from the previous 64,000 to an industry-leading 1 million, enabling deep reasoning and the generation of hundreds of thousands of tokens within a single execution path.

Google Internal Use Cases

Google is actively utilizing Argon in its internal workflows.

• Quantum Algorithm Optimization: Contributed to improving performance by more than 40% compared to existing baselines.
• Memory Efficiency: Secured over 300 TiB after deployment through memory optimization across data centers, with savings of up to 1 PiB expected.
• Code Migration: Working on converting C/C++ codebases to Rust; for 'libgav1', completed a memory-safe video decoder that is 2.7x faster than the existing Rust version.

Benchmark Results

Argon achieved SOTA (state-of-the-art) results on several major benchmarks.

• DeepSWE v1.1: Recorded 77.9%
• LVBench: Recorded 91.7% (long-form video understanding)
• AutomationBench: Recorded 51.3% (end-to-end execution capability)
• Vals Index: Leading in weighted evaluations of economic impact across fields such as finance and law

Cybersecurity and Safety

Argon has been trained to autonomously detect and patch critical software vulnerabilities. It recorded a score of 68% on CWE-bench v1, tying for first place, and took the lead in defensive robustness on Gray Swan's indirect prompt injection (IPI) benchmark. Google has strengthened frontier safety measures to prevent misuse, defend against prompt injection, monitor alignment failures, and enhance system security.