After announcing Gemini 3.5 Pro in June and spending the subsequent months releasing smaller Flash-tier models, Google is returning to the forefront of AI development with Gemini 4 Argon. The company states the model delivers leading-edge capabilities in software engineering, knowledge work, and cybersecurity. However, public access is not currently available.
While external users must wait for a release, internal Google engineers are reportedly already leveraging the new system extensively. According to the company, Argon utilized fleet-wide telemetry data to help reduce memory consumption by 300 TiB across Google’s data centers. Additionally, AI agents built on Argon have assisted in migrating C and C++ codebases to Rust, handling thousands of lines in core libraries such as re2 and libgav1, as well as more than 800,000 lines within the Fuchsia OS Zircon kernel.
To substantiate these claims, Google released a series of performance benchmarks. On the DeepSWE v1.1 software engineering evaluation, Gemini 4 Argon achieved a score of 77.9 percent, outperforming competitors including GPT-6 Astra, Fable 5.1, and Opus 5.5. The company also highlighted strong results in long-horizon tasks, noting an industry-leading score on the Vals Index economic analysis test.
300 TiB saved on memory? Now that’s a benefit for everyone, even if the model itself stays locked away for now.
The cybersecurity claims are what really interest me. Hope they open source the evaluation metrics so we can verify the results.
Meh. Every tech giant promises frontier performance these days. I’ll believe it when I can sign up and test it myself.
Wait, they actually migrated 800k lines of Rust code autonomously? That’s wild. When can we normal users access this?
77.9% on DeepSWE is impressive, but keeping Argon internal feels like classic Google strategy—announce first, deliver later.