Google drops Gemini 4 Argon with a massive 1M output token limit
Google announced Gemini 4 Argon. It generates up to 1 million tokens in a single output and autonomously patches code vulnerabilities.
Models · Source: Hacker News
What happened
Google just revealed Gemini 4 Argon. It is a new frontier model built for long-horizon tasks and deep reasoning. The biggest upgrade is the output token limit. Argon can generate up to one million tokens in a single run. This is a massive jump from the previous limit of 64K tokens. This allows the model to solve complex problems in one continuous trajectory without stopping.
The model is already doing heavy lifting inside Google. It is migrating massive C and C++ codebases to Rust, scaling up to 800,000 lines for the Fuchsia Zircon kernel. One agent rewrote 32,000 lines of video decoder code to run 2.7 times faster than the original port. Another agent found memory optimizations that will save up to one petabyte of memory across Google data centers. Argon even helped quantum computing researchers optimize spacetime resources, beating a published baseline by 40 percent in minutes.
Google is rolling out Argon slowly. Right now, it is only available to trusted cybersecurity defenders through the Fairwind Program. These testers get a version without cyber guardrails to find and patch vulnerabilities. Wiz is already using it and found a critical healthcare vulnerability that older models missed. General access for developers and API customers is coming soon after Google finishes testing safeguards against prompt injections and misuse.
Key facts
- 1M — New output token limit, up from 64K
- 77.9% — Score on DeepSWE v1.1 real-world software engineering benchmark
- $2 — Introductory price per million input tokens
- 95% — Discount applied to cached input tokens
- 2.7x — Speed increase of Rust video decoder rewritten by Argon
Why it matters
The one million token output limit changes how we build autonomous agents. You no longer need to hack together complex loops to generate large codebases or massive documents. The model has the headroom to think deeply and spit out hundreds of thousands of tokens in one go. This makes end-to-end software engineering and deep financial research much more viable. It scored 77.9 percent on the DeepSWE benchmark, proving it can handle real software engineering tasks.
The second-order effect is a shift in cybersecurity dynamics. Google is handing an uncensored version of Argon to defenders. As AI gets better at finding zero-days, the gap between attackers and defenders will shrink or widen depending entirely on who gets access to these raw models first. Google is also actively engaging in the U.S. government voluntary process for pre-release model access. This sets a precedent. Future frontier models will likely require government review before they reach the hands of everyday founders and builders.
For builders
Massive legacy code migrations are now viable
Argon handles up to 800,000 lines of code for migrations. Founders can build tools that port legacy enterprise software to modern memory-safe languages like Rust. Companies stuck on old tech stacks will pay heavily for this automation to avoid manual rewrite costs.
Exploit the cheap cached input pricing
Argon costs two dollars per million input tokens initially, but cached inputs get a 95 percent discount. Builders can load massive codebases or legal libraries into context and query them repeatedly for pennies. This makes heavy-context applications highly profitable for early adopters.
Prepare for agent monitoring requirements
Google monitors Argon chain-of-thought to stop misaligned actions and halts execution when necessary. If you build autonomous agents, you will need similar infrastructure. Enterprise clients will demand proof that your agents cannot go rogue and execute unintended commands on their servers.
My take
Google finally figured out that agents need room to breathe. Capping outputs at 64K tokens was choking real engineering work. By pushing the output to one million tokens and giving defenders an uncensored model, they are actually building for power users instead of just demoing chat bots. The introductory pricing is a clear land grab to lock developers into the Google ecosystem before the price doubles.
Original reporting: Hacker News. This is my rewrite and opinion.