Gemini 4 Argon is Google DeepMind’s newest frontier model, delivering deep reasoning for complex software‑engineering, legal, finance and cybersecurity workflows. It supports an industry‑leading 1 million‑token context window, up from the previous 64 K, allowing a single inference to generate hundreds of thousands of tokens and enabling much deeper long‑horizon problem solving.
Internally, Argon has already accelerated quantum algorithm optimization by 40% in resource‑time trade‑offs, freed over 300 TiB of memory across Google data‑centers through fleet‑wide telemetry analysis, and migrated tens of thousands to over 800 K lines of C/C++ code to Rust in projects such as the Fuchsia Zircon kernel. A Rust port of libgav1 was rewritten with Argon‑guided SIMD experiments, achieving a 2.7× speed‑up while preserving safety.
On benchmark suites Argon leads: DeepSWE v1.1 scores 77.9% on long‑horizon software‑engineering tasks, it tops the Vals Index that weights finance, legal and tax work by U.S. GDP contribution, ranks #1 on AutomationBench with a 51.3% score, and achieves state‑of‑the‑art 91.7% on LVBench video understanding. In security, Argon autonomously discovers, validates and patches vulnerabilities, tying for first on CWE‑bench v1 with a 68% score and helping Wiz’s Scan for Good initiative uncover a critical patient‑data leak in global hospital software.
Google’s safety strategy for Argon covers four pillars: preventing misuse and refusing CBRN‑related requests; hardening against indirect prompt‑injection attacks through red‑team testing and adversarial training; monitoring chain‑of‑thought to stop misalignment; and hardening sandboxed environments while sharing best practices with partners. The model is currently rolling out via the Fairwind program to trusted cyber defenders, priced at $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95%.
Gemini 4 Argon is positioned as a partner for developers, professionals and enterprises to tackle their toughest challenges. Ongoing feedback from early testers will guide further guard‑rail improvements before a broader release to paid API customers and Google AI Ultra subscribers.
Review