Google Gemini 4 Argon to Release With 1 Million Token Limit, 77.9% Coding Score and $2 Input $10 Output Pricing

Google is preparing to release Gemini 4 Argon, a new frontier artificial intelligence model designed for complex, long-horizon workflows across software engineering, enterprise knowledge work and cybersecurity. Google DeepMind has announced the model ahead of its broader availability, with initial access being provided to a limited group of trusted cyber defenders and testers through the Fairwind Program.

According to Google, Gemini 4 Argon will be designed to sustain deeper reasoning across complex tasks that require multiple steps and extended execution. The company said the model is already being tested internally by Google teams for coding, research, engineering and quantum computing-related work.

Gemini 4 Argon to Support 1 Million Output Tokens

One of the key features planned for Gemini 4 Argon will be its expanded output token capacity. Google said the model will support up to 1 million output tokens, compared with the previous 64,000-token limit.

The larger output capacity is expected to allow Gemini 4 Argon to work through lengthy and complex tasks in a single trajectory, reducing the need to divide certain workflows into multiple interactions.

Google said its teams are already testing Argon for large-scale software engineering tasks, including code migrations, algorithm development, debugging and optimization.

Google Reports Strong Software Engineering Performance

Google said Gemini 4 Argon recorded a 77.9% score on the DeepSWE v1.1 benchmark, which evaluates performance on long-horizon software engineering tasks.

The company said Argon will be used for large-scale migration of C and C++ codebases to Rust across Google. These projects range from smaller libraries to larger systems, including the Fuchsia Zircon kernel.

In one example involving libgav1, Google said Argon agents replaced around 32,000 lines of SIMD code in an existing Rust implementation through repeated profile-guided experiments and compiler analysis.

According to Google, the resulting Rust-based video decoder runs 2.7 times faster than the previous Rust port while producing identical video output.

Google said large-scale code migrations will undergo automated and manual auditing, emulation testing and reviews before being deployed to production environments.

Gemini 4 Argon to Target Enterprise AI Workflows

Beyond software development, Gemini 4 Argon will be aimed at enterprise applications across finance, legal work, research and other knowledge-intensive areas.

Google reported a 68.9% score for Argon on the Vals Index and 65.4% on Vals Finance Agent v2. The company also reported scores of 19.6% on the Harvey’s Legal Agent Benchmark and 51.3% on AutomationBench.

The model will also support tasks involving visual and multimodal information. Google reported a 91.7% score on LVBench, a benchmark focused on long-video understanding.

These capabilities are expected to support workflows involving documents, charts, videos and other forms of business information.

Gemini 4 Argon to Focus on Cybersecurity Defense

Cybersecurity will be another major focus of Gemini 4 Argon. Google said the model will be designed to help cyber defenders identify, validate and patch software vulnerabilities.

Google reported that Argon achieved 68% on CWE-bench v1, tying for the highest score on the benchmark. The company also said the model performed strongly in internal vulnerability assessments and penetration-testing evaluations.

Through the Fairwind Program, Google is initially providing access to a selected group of trusted cyber defenders. The early deployment will allow the company to collect real-world feedback while continuing to strengthen the model’s safety measures.

Google to Strengthen AI Safety Measures Before Wider Release

Google said the broader release of Gemini 4 Argon will follow a phased approach because of the model’s advanced capabilities.

The company said it will continue strengthening safeguards against potential misuse, including risks involving cyberattacks and chemical, biological, radiological and nuclear applications. Google also said internal and external red-team testing will be used to evaluate the robustness of its safeguards.

Additional safety measures will address indirect prompt injection attacks, potential model misalignment and the security of sandboxed environments used for high-risk training and evaluations.

Google said it is also participating in the U.S. government’s voluntary process for pre-release access to advanced AI models.

Gemini 4 Argon Pricing and Availability

Google said Gemini 4 Argon will launch at an introductory price of $2 per million input tokens and $10 per million output tokens. Cached input tokens will be priced at 95% below the standard input-token price.

The model is currently being introduced to a limited group of trusted testers and cyber defenders. Google said it will make Gemini 4 Argon available more broadly as soon as possible, with the rollout expected to reach developers, enterprises and consumers.

The company said broader availability will begin with paid API customers and Google AI Ultra subscribers.

Google said feedback from the initial group of cyber defenders and trusted testers will be used to improve the model and strengthen its safeguards before wider release.