SAN FRANCISCO, Oct 1 (Zark News) — Google on Wednesday unveiled Gemini 4 Argon, which the company calls its most powerful AI model yet, with cybersecurity as its signature capability — and is limiting initial access to a select group of trusted cyber defenders.
Argon was trained specifically for defensive cyber work and can “autonomously find, validate, and patch critical software vulnerabilities,” according to Google’s announcement, as reported by TechCrunch. The model is being rolled out first through the company’s Fairwind Program, its security initiative, rather than as a general release. Google says the phased approach lets early users test the model, identify vulnerabilities and improve safeguards before broader availability.
The model is built for long, multi-stage workflows: its output limit jumps from 64,000 tokens to 1 million, allowing it to work through lengthy coding projects and business workflows in a single run. Google says thousands of its own employees already use Argon for debugging and codebase migrations, including migrating more than 800,000 lines of the Fuchsia Zircon kernel from C and C++ to the memory-safe Rust language, according to technology outlets.
Google claims striking internal results: Argon helped quantum-computing researchers improve an algorithm’s efficiency by 40% over a published baseline, and Argon-powered agents analyzing data-center telemetry freed more than 300 terabytes of memory, with total savings eventually reaching up to a petabyte. In another project, Argon made a Rust version of Google’s libgav1 video decoder 2.7 times faster than the earlier Rust port, according to the company.
On benchmarks — figures cited by Google and not yet independently confirmed — Argon scored 77.9% on DeepSWE v1.1 for long-horizon software engineering, 51.3% on Zapier’s AutomationBench, 91.7% on LVBench for long-video understanding, and 68% on CWE-bench for fixing software vulnerabilities. Google claims Argon significantly outperformed OpenAI’s GPT-6 Astra and Anthropic’s Fable and Opus models across benchmarks, citing the AI benchmarking startup Vals.
The restricted rollout marks a notable shift in how frontier labs deploy powerful models, the Digital Watch Observatory noted: by placing cyber defenders first in line, Google is effectively turning early deployment into an additional security-testing phase. Introductory pricing is $2 per million input tokens and $10 per million output tokens, with cached inputs discounted 95%.
Sources: TechCrunch, Digital Watch Observatory
