Tech Desk: Google has just unveiled Google Gemini 4 Argon, its most formidable AI model to date, setting a new benchmark in the competitive AI landscape. This cutting-edge model, the first in the Gemini 4 series, is engineered for complex, multi-step tasks across software engineering, legal, finance, and cybersecurity defense, directly challenging rivals like OpenAI’s GPT-6 Astra and Anthropic’s Claude Opus 5.5.

Initial benchmarks reveal Google Gemini 4 Argon’s superior performance in critical areas. It scored 77.9% on DeepSWE v1.1 for software engineering, surpassing both Claude Opus 5.5 and GPT-6 Astra. The model also demonstrated a significant lead on Harvey’s Legal Agent Benchmark, hitting 19.6% compared to Astra’s 5.4%. Notably, Google has dramatically increased its output limit to 1 million tokens, enabling Argon to complete extensive tasks in a single run, and offers competitive API pricing.

However, access to Google Gemini 4 Argon is currently limited. It’s first being rolled out to a select group of cyber defenders via Google’s Fairwind Program, followed by paid API customers and Google AI Ultra subscribers. CEO Sundar Pichai emphasized a cautious approach, stating availability would be “as soon as we can and as safely as we can.”

Google is already leveraging Argon internally, with agents freeing up substantial data center memory and migrating C/C++ codebases to Rust. Significant safety guardrails are also in place, training the model to refuse requests aiding cyberattacks or CBRN weapons, and enhancing its resistance to indirect prompt injection. This strategic rollout positions Google Gemini 4 Argon as a powerful contender, signaling rapid advancements in AI capabilities.