Gemini 4 Argon

Google has announced Gemini 4 Argon, its new frontier AI model built for deep reasoning across complex, long-horizon workflows. Described as a major step in frontier intelligence, Argon targets three core areas: real-world software engineering, enterprise knowledge work in legal and finance, and cybersecurity defense.

Unlike a general chatbot upgrade, Argon is designed to sustain reasoning over hundreds of thousands of tokens in a single run. To enable this, Google expanded its output limit to an industry-leading 1 million tokens, up from 64K. That headroom lets the model think deeply, plan multi-step work, and solve tough problems in one go rather than in fragmented pieces.

Argon is already changing how Google works internally. Thousands of Googlers are using it for specialized coding, deeper research, and higher-quality writing. Standout examples include:

• Quantum optimization: Argon beat a published baseline by 40% in minutes when optimizing spacetime resources (qubits x gates) for quantum subroutines.
• Memory efficiency: Teams of Argon agents analyzed fleet-wide telemetry to find memory savings across Google's data centers, freeing over 300 TiB so far, with an estimated 500 TiB to 1 PiB total.
• C++ to Rust migrations: Argon agents are migrating critical codebases from tens of thousands of lines in libraries like re2 and libgav1 up to 800K+ lines for the Fuchsia OS Zircon kernel. For libgav1, it rewrote 32K lines of SIMD code into safe Rust that the compiler auto-vectorizes, running 2.7x faster with identical video output.

On benchmarks, Argon sets new highs for practical work. It leads DeepSWE v1.1 for long-horizon software engineering at 77.9%, tops the Vals Index for GDP-weighted economic impact across finance, coding, legal and tax, and ranks #1 on Zapier's AutomationBench at 51.3%. It also leads on Vals Finance Agent v2, Harvey's Legal Agent Benchmark, and LVBench for long-video understanding at 91.7%.

Cybersecurity is a defining focus. Argon can autonomously find, validate, and patch critical vulnerabilities. Wiz is already using it in its Scan for Good initiative, where it found a critical healthcare vulnerability missed by previous frontier models. It ties for first on CWE-bench v1 at 68% for remediation.

For now, Google is taking a phased rollout. Argon is going first to trusted cyber defenders via the Fairwind Program (without cyber guardrails for defense use) while participating in the U.S. government's voluntary pre-release process. Google cites four safeguard areas before broad release: misuse prevention for cyber and CBRN risks, prompt-injection resilience, misalignment monitoring, and hardened sandboxes. Pricing starts at $2 per million input tokens and $10 per million output tokens, with broader access for paid API customers and Google AI Ultra subscribers coming soon.

Source: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon/