Google DeepMind Unveils Gemini 4 Argon with 1M Output Tokens for Coding, Knowledge Work and Cyber Defense
MarkTechPost Asif Razzaq ● Covered by 7 sources
Google DeepMind unveiled Gemini 4 Argon with 1M-token outputs. It’s aimed at coding and cyber defense, but access is limited for now.
Based on reporting by MarkTechPost, Asif Razzaq — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Google DeepMind has introduced Gemini 4 Argon, the first model in its Gemini 4 line. It’s being pitched as a frontier model for long software projects, enterprise knowledge work in legal and finance, and cybersecurity defense. The standout change is simple and huge: it can produce up to 1 million tokens in a single reply, where earlier Gemini models topped out at 64K.
That output ceiling is the real story here. Google is betting that some jobs are no longer about getting a model to answer well, but about keeping it on task for a very long stretch. Large refactors, long reports, and chained analysis can stay inside one run instead of being chopped into pieces. The price tag is equally explicit: introductory pricing is $2 per 1M input tokens and $10 per 1M output tokens, with cached input discounted to $0.10 per 1M. After the intro period, it rises to $4 and $20.
On Google’s own benchmark table, Argon comes out ahead on 12 of 18 tests and ties for first on one more. It scores 77.9% on DeepSWE v1.1, 68.9% on Vals Index, 51.3% on AutomationBench and 19.6% on Harvey Legal Agent Benchmark. It also posts 91.7% on LVBench, which Google says is a new state of the art. But the model is not a clean sweep. It trails GPT-6 Astra on FrontierSWE v2 and OSWorld-2.0, and it trails Claude Opus 5.5 on Terminal-Bench 4.0.
Google says it is taking a staged rollout. Argon is in the U.S. government’s voluntary pre-release access process, with feedback from early testers feeding into guardrails before a wider release. And this is not just a coding toy. Google says trusted defenders and internal teams can use it without cyber guardrails, while broader release is gated by defenses for misuse, indirect prompt injection, misalignment monitoring and isolated sandboxes. That’s a very Google way to say: yes, it can do a lot, and yes, we’re going to be careful about who gets the sharp edges.
My take — AI-written commentary, not fact-checked reporting
This is the part of AI that actually matters: long, expensive, high-trust work, not another chatbot demo with a fresh coat of paint. The 1M-token number is loud, but the tighter access and extra safety work matter more. Everyone else is still selling scale; Google is selling control, which is usually the more honest business.
Read more about this at: MarkTechPost