Google Unveils Gemini 4 Argon Flagship AI Model

The Tech TL;DR: Gemini 4 Argon

• Pricing and Scaling: Priced at $2 per million input tokens and $10 per million output tokens, with cached input discounted by 95%. Trajectory output capacity scales up to 1 million tokens.
• Enterprise Benchmarks: Scores 68% on CWE-bench v1 for cybersecurity vulnerabilities, tying for top performance while handling automated patching.
• Deployment Constraints: Initial access is strictly limited to select cyber defenders and internal Google teams during safety testing, with broader availability scheduled for paid API customers and Google AI Ultra subscribers.

Google Restricts Gemini 4 Argon to Select Users

Access to Gemini 4 Argon is currently restricted to select cyber defenders and Google’s internal engineering teams while the company runs rigorous safety evaluations. Tulsee Doshi, Google’s Gemini model product lead, described the system as an incredibly well-rounded architecture capable of handling multi-step tasks. Broader availability will follow in phases, starting with paid API customers and Google AI Ultra subscribers, though Google has not provided a public release date.

Concurrently, Google is participating in the Trump administration’s voluntary pre-release model access process. This rollout follows a White House AI safety agreement signed by Google CEO Sundar Pichai alongside President Donald Trump and other technology executives.

Performance Metrics and Architectural Pricing Comparison

Gemini 4 Argon expands trajectory output capacity significantly, enabling the generation of up to 1 million output tokens in a single run, compared to 64,000 tokens in previous iterations. Financial services firm Jefferies noted that Argon establishes a new benchmark for the cost-performance frontier in foundation models, pricing tokens well below competitors like Anthropic’s Claude Fable 5.1 and Claude Opus 5.5.

On enterprise benchmarks, Google stated that Argon competes directly with OpenAI’s Astra and Anthropic’s Opus 5.5. The model also ranks ahead of competing systems on the VALS Index and ties with OpenAI’s GPT-6 Astra and Grok 4.7 on specific cybersecurity evaluations.

Internal Infrastructure Optimization and Code Migrations

Thousands of Google employees are already utilizing the model in production environments. Argon-driven internal agents successfully identified memory optimizations that reclaimed more than 300 terabytes of operational memory across data centers, with total projected savings estimated between 500 terabytes and 1 petabyte.

Beyond data center resource allocation, the model optimized a quantum computing subroutine by 40%. Engineering teams are also deploying Argon to automate large-scale C/C++ to Rust migrations, rewriting more than 800,000 lines of code for the Fuchsia Zircon kernel.

Cybersecurity Evaluation and Vulnerability Patching

In automated security assessments, Gemini 4 Argon achieved a score of 68% on CWE-bench v1, tying for the highest recorded score. Google stated that the model can autonomously discover, validate, and patch software vulnerabilities. Trusted cybersecurity partners are being granted access without standard safety guardrails to evaluate its full defensive capabilities.