Google launches Gemini 4 Argon flagship model

Google unveiled Gemini 4 Argon, its new flagship model with up to 1 million tokens per output, priced at GPT-6.1 Sol level but competing with Fable. It leads in benchmarks like DeepSWE (77.9%) and Harvey legal (19.6%). Initial access limited to trusted testers and cybersecurity personnel.
Key insights
Massive Output Capacity
Gemini 4 Argon can output up to 1 million tokens in a single run, nearly 8 times the 128k limit of GPT-6 Astra and Opus 5.5. This allows continuous deep reasoning without splitting tasks, enabling analysis of very long documents.
Competitive Pricing
During promotional period, Argon costs $2/M input tokens and $10/M output, matching GPT-6.1 Sol and Claude Opus 5.5. After promotion, rises to $4/$20, still one-fifth the price of Astra and Fable. Cached input charged at 5% of base.
Benchmark Leadership
Argon leads 13 of 19 benchmarks: DeepSWE v1.1 77.9%, Harvey legal 19.6% (vs 5-7% for opponents), Vals Finance Agent 65.4%, Vals Index 68.9%, Zapier AutomationBench 51.3%. On long video LVBench 91.7%. However, it trails on FrontierSWE v2, Terminal-bench 4.0, and OSWorld-2.0.
Internal Google Success Stories
Argon agents optimized quantum algorithm resource consumption 40% better than published baseline, released over 300 TiB of memory across data centers, and completed migration of 800k+ lines of C/C++ code to Rust, with 2.7x speedup in libgav1 Rust port.
Cybersecurity Focus
Argon is specifically trained for cyber defense: finding, verifying, and patching vulnerabilities. On CWE-bench v1 it ties Astra at 68%. It is initially released to Google's Fairwind program for trusted defenders. Wiz Scan for Good already uses Argon, discovering a critical vulnerability in hospital software.
Safety Measures
Google implements four safety measures: refusing harmful requests with activation monitoring, anti-prompt injection, monitoring thinking and actions for deviance, and isolating sandbox environments for high-risk testing. Argon undergoes US government pre-release voluntary review.
Gradual Rollout Plan
Argon is currently available only to trusted testers and cybersecurity defenders in Fairwind. Next, paid API and Google AI Ultra subscribers will get access, then consumers. Specific dates unannounced. The model is also under US government pre-release review.