Google unveils Gemini 4 Argon, a frontier model for long-horizon reasoning and cybersecurity defense

Google announced a new frontier model, Gemini 4 Argon, designed for sustained deep reasoning across complex, long-horizon workflows. The company is initially releasing the model to trusted cyber defenders through its Fairwind Program, while gradually expanding access and collecting feedback from early testers.

Capabilities and internal use

According to Google, Argon is already powering internal workflows: thousands of Googlers use it for specialized coding tasks, deeper research, and writing. Notable results cited in the announcement include:

• Quantum algorithm optimization: Argon reportedly beat a published baseline by 40% in minutes when optimizing spacetime resources (qubits × gates) for bottleneck subroutines.
• Memory efficiency: Argon agents analyzed fleet-wide profiling telemetry and autonomously applied memory optimizations in Google data centers, freeing over 300 TiB of memory after rollout, with total savings estimated between 500 TiB and 1 PiB.
• Large-scale codebase migrations: Argon agents are assisting C/C++ to Rust migrations across Google, ranging from tens of thousands of lines in core libraries like re2 and libgav1 up to more than 800,000 lines in the Fuchsia OS Zircon kernel. These critical rewrites undergo rigorous automated and manual auditing, emulation testing, and review before production rollout.

For libgav1, Google’s open-source video decoder, Argon took an existing Rust port and replaced 32K lines of SIMD code through many rounds of profile-guided experiments and compiler-output analysis, producing safe Rust that the compiler would vectorize. The resulting memory-safe decoder runs 2.7× faster than the Rust port while producing identical video output, approaching the performance of optimized C++.

Extended context window: 1 million token limit

To support deeper, longer tasks, Argon’s output token limit has been increased from 64K to an industry-leading 1,000,000 tokens. Google says the extra headroom allows the model to reason more deeply and generate hundreds of thousands of tokens in a single trajectory, enabling more complete solutions to complex problems in one pass.

Performance across enterprise and professional workflows

Google reports leading performance for Argon on multiple benchmarks and domain-specific evaluations:

• DeepSWE v1.1 (long-horizon software engineering): 77.9%.
• Vals Index (economic impact across finance, coding, legal, and tax, weighted by U.S. GDP): Argon ranks as the leading model.
• Other domain benchmarks where Argon leads include Vals Finance Agent v2 and Harvey’s Legal Agent Benchmark.
• AutomationBench (Zapier’s end-to-end business function benchmark): Argon ranks #1 with a score of 51.3%.
• LVBench (long video understanding): Argon reportedly scores 91.7% and leads the benchmark.

Argon’s multimodal strengths are highlighted for tasks that require visual understanding, such as professional chart analysis, extracting details from long videos, and acting on sequences of documents.

Cybersecurity defense and vulnerability discovery

Gemini 4 Argon was trained for cybersecurity defense and can autonomously find, validate, and patch critical software vulnerabilities. For trusted defenders and Google’s internal teams, Argon will be available without cyber guardrails so they can use its full frontier-level defensive capabilities.

• Wiz is using Argon in its Scan for Good initiative, which aims to protect critical public infrastructure for free. In an early demo, Argon uncovered a critical vulnerability in healthcare software exposing sensitive personal data that earlier frontier models had missed.
• On CWE-bench v1 (measuring remediation capability), Argon ties for first with a top score of 68%.
• On Google’s internal comprehensive vulnerability benchmark, Argon identified a wide range of exposures across complex codebases in 20 programming languages.
• On Wiz’s internal black-box penetration testing benchmark, Argon outperformed 3.8 Flash Cyber in discovering attack surface, identifying vulnerabilities, and producing proof-of-concept evidence.

Strengthening safeguards and phased release

Before broad availability, Google is enhancing frontier safeguards in four areas:

1. Defending against misuse: Argon is designed, under Google’s Frontier Safety Framework, to refuse harmful requests related to cyberattacks or CBRN topics while preserving legitimate dual-use scientific research. Google reports robustness testing of these safeguards by internal and external red teams.

2. Protecting against prompt injection: Argon is presented as Google’s most resilient model to date against indirect prompt injection attacks. The company used automated red teaming and adversarial training and reports leading results on the Gray Swan Indirect Prompt Injection benchmark.

3. Monitoring for misalignment: Google deploys mitigations that monitor Argon’s chain-of-thought and actions, stopping execution when necessary to prevent the model from exceeding the user’s intentions. A similar monitoring system was used during training runs to alert a dedicated incident response team; Google says it avoids feeding such findings back into training to prevent shaping the model’s reasoning to evade detection.

4. Hardening systems: To safely test frontier models, Google said it isolates and seals sandboxed environments before high-risk training or evaluations, aligned with its agent control roadmap, and plans to share agent security best practices with partners.

Pricing and availability

Introductory API pricing for Argon is set at $2 per 1M input tokens and $10 per 1M output tokens; cached input tokens are priced at 95% off the input token price. After the introductory period, prices will rise to $4 per 1M input tokens and $20 per 1M output tokens.

Argon will first be available to the initial cohort of cyber defenders and trusted testers in the Fairwind Program, and Google plans to expand access to developers, enterprises, and consumers—starting with paid API customers and Google AI Ultra subscribers—after iterating on safeguards and gathering feedback.

Conclusion

Google positions Gemini 4 Argon as a frontier model that advances long-horizon reasoning, coding productivity, enterprise knowledge work, and cybersecurity defense. The company emphasizes a phased, safety-focused rollout backed by red-teaming, monitoring of internal activations and chain-of-thought, and hardened sandbox environments before broader distribution.