Google launches Gemini 4 Argon, billed as its strongest model to date

Google describes Argon as its most capable model to date, built to sustain deep reasoning across complex, long-running workflows. Rather than a general-purpose chatbot upgrade, the company is pitching it squarely at real-world software engineering, enterprise knowledge work in legal and financial domains, and cybersecurity defense.

The launch also signals a strategic reset. Google has cancelled its previously planned Gemini 3.5 Pro and is putting its full weight behind the new Gemini 4 generation, with Argon as its first release.

Cybersecurity Defenders Get First Access

The rollout is deliberately narrow. Initial access is being granted exclusively to a set of trusted cyber defenders through Google's Fairwind Program. These vetted users will receive access without the usual cyber guardrails, allowing them to exercise the model's full capabilities.

Google is also participating in the U.S. government's voluntary pre-release model access process. Broader availability for paid API customers and Google AI Ultra subscribers will follow later, though the company has not announced a specific date.

Technical Specifications

Two numbers stand out in the spec sheet:

- Output token limit: Raised from 64,000 to 1 million tokens, enabling the model to generate enormous volumes of code or sustain very long reasoning chains in a single run.

- Pricing: An introductory rate of $2 per million input tokens and $10 per million output tokens, with cached inputs discounted by 95%. Prices are expected to double once the introductory period ends.

Built to Find and Fix Vulnerabilities

Google claims Argon can autonomously identify, validate, and patch critical software vulnerabilities. In early testing, the company says the model discovered a critical flaw in healthcare software that earlier frontier models had missed.

On CWE-bench v1, which measures vulnerability remediation, Argon scored 68%, tying for first place.

Benchmark Claims and Caveats

Google published a comparison table pitting Argon against OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5, claiming leadership in 13 of 19 benchmarks. Highlights include:

Benchmark What It Measures Argon Score

DeepSWE v1.1 Long-horizon software engineering 77.9%

Vals Index Knowledge work 68.9%

AutomationBench Zapier business functions 51.3%

LVBench Long video understanding 91.7%

Those figures come with important context. On the independent Artificial Analysis Intelligence Index, Argon does not hold the top spot. It ranks behind Claude Opus 5.5 and roughly level with GPT-6 Astra. Some internal Google staff have also reportedly expressed skepticism about how the model performs in practice compared with its benchmark scores.

Already Working Inside Google

Google says thousands of its own employees are using Argon for internal workflows. Notable examples include:

- Codebase migration: Argon agents are helping migrate massive C/C++ codebases to Rust, including the 800,000+ line Fuchsia OS Zircon kernel and core libraries such as re2 and libgav1.

- Memory optimization: Argon agents autonomously identified optimizations that freed more than 300 TiB of memory across Google's data centers.

- Quantum algorithms: The model helped optimize quantum computing subroutines, exceeding published baselines by 40% within minutes.

What to Watch

Argon's success will hinge on whether its benchmark dominance translates into dependable performance in the hands of outside developers and security teams, and on how quickly Google widens access beyond the Fairwind Program. For now, the model is a statement of intent: Google wants to be taken seriously as the vendor of choice for the hardest, most technical work in AI.