Google Drops Gemini 4 Argon: 1M Output Tokens and Uncensored Cyber Capabilities
Google just announced Gemini 4 Argon. It writes 1 million tokens at a time, rewrites C++ to Rust, and ships to cyber defenders first.
Models · Source: Hacker News
What happened
Google just revealed Gemini 4 Argon. It is their newest frontier model built for long-horizon tasks. The standout feature is a massive 1 million token output limit. This gives the model headroom to generate hundreds of thousands of tokens in a single run. It is designed for deep reasoning across coding, finance, and legal workflows. Google wants this model to solve tough problems in one go rather than forcing developers to chain together small outputs.
Google is already using Argon internally with wild results. Argon agents are migrating legacy C and C++ codebases to Rust across the company. In one case, it rewrote a video decoder library called libgav1. The new memory-safe Rust version runs 2.7 times faster than the previous port. Argon also autonomously found memory optimizations across Google data centers. This freed up over 300 TiB of memory, with total savings estimated to reach 1 PiB.
The rollout is heavily restricted for now. Google is launching Argon first to trusted cybersecurity defenders through its Fairwind Program. These defenders get a version without cyber guardrails so they can use its full defensive capabilities. Wiz is already using it to scan public infrastructure. They found a critical vulnerability in healthcare software that older models missed. A public API release for developers and enterprise customers is coming soon.
Key facts
- 1M — Output token limit, up from 64K
- $2 — Introductory price per million input tokens
- $10 — Introductory price per million output tokens
- 95% — Discount for cached input tokens
- 77.9% — Score on DeepSWE v1.1 software engineering benchmark
- 300 TiB — Memory freed up in Google data centers by Argon agents
- 2.7x — Speed increase of libgav1 video decoder after Argon rewrote it in Rust
- 68% — Top score on CWE-bench v1 for remediating security vulnerabilities
Why it matters
The 1 million output token limit changes how we build agentic workflows. You no longer need to stitch together hundreds of small API calls to generate a massive codebase or a comprehensive legal brief. You can ask Argon to solve a complex problem and let it think and write in one continuous trajectory. This breaks the output bottleneck that has choked complex AI agents for years. Builders can finally design systems that output entire applications or exhaustive research reports in a single shot.
Releasing an uncensored cyber model to trusted partners signals a massive shift in AI safety strategy. Google is acknowledging that top-tier defensive security requires models that know exactly how to attack. By giving companies like Wiz unrestricted access, Google is weaponizing AI for defense before bad actors can exploit these capabilities. This will force other AI labs to rethink their rigid safety filters. You cannot build enterprise security tools if your AI refuses to write an exploit payload.
For builders
Massive cost savings with prompt caching
Argon launches at $2 per million input tokens, but cached inputs get a 95 percent discount. Builders who optimize their context windows will pay pennies for massive prompts. Start structuring your applications to reuse context heavily across multiple user sessions.
Legacy code migration is automated
Google is using Argon to rewrite hundreds of thousands of lines of C and C++ into memory-safe Rust. If you run a development agency, selling automated legacy code migration just became a highly profitable business model. The model does the heavy lifting while your engineers handle the auditing.
Output tokens are very expensive
The introductory output price is $10 per million tokens, jumping to $20 later. Generating a full 1 million token response will cost up to $20 per API call. You must strictly control when your application triggers these massive generations or your cloud bill will destroy your margins.
My take
Google finally figured out that output limits were holding AI agents back. Giving a model a massive context window is useless if it can only spit out a few pages of text. I love that they are letting trusted security teams use the model without guardrails. We need AI that actually works in the real world, not lobotomized chatbots that refuse to do their jobs.
Original reporting: Hacker News. This is my rewrite and opinion.