Gemini 4 Argon Operational

Gemini 4 Argon is a new-generation frontier model developed by Google DeepMind, designed for complex, long-running workflows across enterprise knowledge work such as software engineering, finance, and law, as well as cybersecurity defense. Gemini 4 Argon increases the maximum single-output length from 64,000 Tokens to 1 million, enabling it to complete long-running tasks such as large-scale code migrations in one pass. Gemini 4 Argon achieved 77.9% on the DeepSWE v1.1 software engineering benchmark and is being released in phases through the Fairwind program, initially for trusted cybersecurity defenders.

Extremely long outputs: The single-output limit has increased from 64,000 Tokens to 1 million, enabling the generation of hundreds of thousands of Tokens of code or in-depth reports in one pass without breaking the context.
Software engineering: Achieves 77.9% on the real-world, long-running DeepSWE v1.1 engineering benchmark, supporting code debugging, algorithm design, and large-scale codebase migrations.
Enterprise knowledge work: Delivers leading performance in finance, law, and tax, ranking first on the US GDP-weighted Vals Index.
Multimodal understanding: Analyzes specialized charts, understands long videos, and takes action based on multiple documents.
Cybersecurity defense: Autonomously discovers, validates, and fixes critical vulnerabilities, tying for first place on CWE-bench v1 at 68% and identifying attack surfaces in black-box penetration tests without source code.
Automated business process execution: Ranks first on Zapier AutomationBench, an end-to-end business execution benchmark.
Safety protections: Includes a four-layer protection system covering abuse prevention, resistance to indirect prompt injection, chain-of-thought misalignment monitoring, and sandbox hardening.

Long-horizon reasoning architecture: Google has increased Argon's output limit from 64K to 1 million Tokens. When the model can generate hundreds of thousands of Tokens within a single task trajectory, it can perform deeper reasoning without interruption and complete complex, multi-step tasks such as large-scale code migrations and in-depth research in one pass.
Agentic engineering capabilities: The model operates through multi-agent collaboration, analyzing telemetry and autonomously implementing memory optimizations, rewriting code through multiple rounds of performance experiments, and more. All large-scale code migrations must pass automated audits, simulation testing, and human review before deployment.
Abuse and injection prevention: Under the Frontier Safety Framework, the model refuses requests involving cyberattacks and CBRN misuse while preserving legitimate dual-use research. Internal activation monitoring and red-team testing improve robustness, while automated red teaming and adversarial training defend against indirect prompt injection.
Misalignment monitoring and system hardening: Argon's chain of thought and actions are monitored throughout execution and halted when they deviate from the user's intent. Similar monitoring systems operate during training and alert the incident response team, while monitoring results are deliberately excluded from training to prevent evasion. Before high-risk training and evaluation, sandbox environments are isolated and sealed.

Confirm eligibility: Gemini 4 Argon is not yet generally available. First determine whether you belong to one of the initial user groups: trusted cybersecurity defenders, paid API customers, or Google AI Ultra subscribers.
Defender application channel: Cybersecurity professionals can apply for early access through the Google Fairwind Program website https://deepmind.google/fairwind-program/. Once approved, they will receive the complete defensive version without network guardrails.
Subscription for general users: Developers and enterprise users must wait for the official phased rollout. Early access may be available through the paid API, while AI Ultra subscribers will receive priority access.
API access: After receiving eligibility, create a project on the Google AI developer platform and call the Gemini 4 Argon API.

Extremely long outputs: The single-output limit reaches 1 million Tokens, nearly 8 times that of most frontier models, which are around 128,000 Tokens, enabling uninterrupted long-running tasks in one pass.
Leading software engineering performance: Achieves 77.9% on the real-world, long-running DeepSWE v1.1 engineering benchmark, setting a new record and supporting Google's migration of hundreds of thousands of lines of C/C++ to Rust internally.
Comprehensive leadership in enterprise knowledge work: Ranks first across finance, programming, law, and tax on the US GDP-weighted Vals Index.
Strong multimodal understanding: Scores 91.7% on the long-video understanding benchmark LVBench, analyzes specialized charts, and takes action based on multiple documents.
First in business automation: Ranks first on the end-to-end business execution benchmark Zapier AutomationBench with 51.3%.
Top-tier cyber defense: Ties for first place on CWE-bench v1 at 68% and has discovered high-severity vulnerabilities affecting medical software used by hospitals worldwide.

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

Codebase migration: Automatically migrates more than 800,000 lines of C/C++ code from core libraries such as Google re2 and libgav1, as well as the Fuchsia Zircon kernel, to memory-safe Rust, with deployment following audits.
Performance-critical code rewriting: In the open-source video decoder libgav1, automatically rewrites 32,000 lines of SIMD code into safe Rust that can be automatically vectorized by the compiler, producing completely identical video output while achieving a 2.7x speedup.
Data center cost reduction: Analyzes performance telemetry across the entire cluster, autonomously identifies and applies memory optimization strategies, and releases more than 300 TiB of memory after deployment, with estimated total savings of 500 TiB to 1 PiB.
Quantum algorithm optimization: Optimizes the product of the number of qubits and gate operations in quantum computing subroutines, exceeding a published baseline by 40% within minutes in one case.