Phased Launch Targets Trusted Partners

Google initiated a controlled rollout of Gemini 4 Argon on September 30, 2026, beginning with a cohort of vetted cybersecurity partners through the Fairwind program. This marks the first public appearance of the flagship model, developed internally and designed for complex technical tasks in network defense and code generation.

The release follows months of internal testing and comes as Google scales back on smaller flash models to focus on frontier capabilities. Access remains restricted initially, with broader availability planned for paid API users and Google AI Ultra subscribers in subsequent phases.

Google also conducted a pre-release safety evaluation alongside the U.S. government, strengthening safeguards against potential misuse before wider distribution.

Performance Benchmarks Set New Standards

Argon recorded a score of 77.9 percent on the DeepSWE v1.1 benchmark, which evaluates long-horizon software engineering tasks in realistic scenarios. This result surpasses previous frontier models and positions the system as a leader in handling extended workflows that once required multiple sessions.

The model ties for first place on the CWE-bench v1 evaluation for vulnerability remediation, achieving a score of 68 percent. It also outperforms earlier versions of Google’s cyber-focused models on internal tests, including those conducted by security firm Wiz.

Additional strengths appear in enterprise knowledge work, with top rankings on the Vals Index that weights contributions to U.S. GDP across finance, legal, and technical fields. Argon also led on automation and video comprehension benchmarks.

Expanded Capacity and Real-World Applications

Argon supports output up to 1 million tokens in a single response, a significant increase from the prior limit of 64,000 tokens. This change enables deeper analysis on complex projects such as code reviews, vulnerability assessments, and multi-stage research without breaking tasks into separate calls.

Inside Google’s infrastructure, the model has already optimized data center memory usage, freeing more than 300 terabytes once fully deployed and an estimated 500 terabytes to 1 petabyte overall. Quantum computing teams have also applied Argon to improve qubit and gate efficiency in subroutines.

Security firm Wiz integrated the model through its Scan for Good initiative and identified a critical flaw in worldwide hospital software that escaped detection by earlier frontier systems. The vulnerability affected sensitive personal information across healthcare platforms.

Pricing and Future Rollout

API access for eligible users carries an introductory rate of $2 per million input tokens and $10 per million output tokens. Standard rates are expected to rise to $4 and $20 respectively once the introductory period concludes.

Google outlined a phased approach to distribution, starting with cybersecurity defenders and trusted testers whose feedback will refine the system before it reaches general developers and enterprises. The model includes built-in protections that refuse requests related to chemical, biological, radiological, or nuclear misuse while still allowing legitimate dual-use research.

Argon also demonstrates resilience against indirect prompt injection attacks, scoring ahead of competitors on Gray Swan’s benchmark. This layered safety approach addresses concerns about the same capabilities that aid defense from also enabling misuse.

Context Within Competitive Landscape

The launch arrives amid a broader push by major U.S. technology companies to demonstrate responsible AI development. It follows Google CEO Sundar Pichai’s signing of a voluntary AI safety accord with President Donald Trump just the previous day.

Argon ties or leads in several cybersecurity evaluations alongside models from OpenAI and Anthropic, while trailing in a few areas such as certain science and control tests. Google maintains the results reflect real-world utility in enterprise and defense settings rather than isolated lab scores.

Analysts note the timing reflects Google’s strategy shift back toward high-capability models after focusing on faster, cost-efficient versions earlier in the year.