Google has announced its new frontier AI model, the Gemini 4 Argon. The model is rolling out to a set of the company’s trusted cybersecurity partners.
According to benchmark results, the model outperforms OpenAI’s GPT-6 Astra, Claude Fable 5.1, and Opus 5.5 on several coding, science, computer use and knowledge work evaluations.
The company shared the model’s performance on numerous internal tasks. The model helped save over 300 TiB of memory in the company’s data centres by applying various optimisations.
In addition, Argon’s agents are helping the company migrate C/C++ codebases to Rust.
In one completed libgav1 optimisation, Argon started with an existing Rust port of Google’s open-source video decoder and replaced 32,000 lines of SIMD code.
The agents repeatedly ran profile-guided experiments, examined compiler output and modified the Rust implementation so that the compiler could automatically vectorise it. Google said the resulting memory-safe decoder runs 2.7 times faster than the previous Rust implementation while producing identical video output.
Google plans to expand access to paid API customers and Google AI Ultra subscribers before a broader rollout. Introductory pricing is set at $2 per million input tokens and $10 per million output tokens.
Gemini 4 Argon ends a long wait for Google, after OpenAI and Anthropic pulled ahead in the frontier AI model race in recent months.
However, reports already show the model struggling. Bloomberg reported on the day of the launch that some Google employees with direct access to Gemini 4 believe its benchmark performance does not fully translate to real-world use.
Sources told the media outlet that the model struggles with certain coding tasks when employees use it, despite performing strongly on industry benchmarks.
Google disputed that characterisation, while another employee cited by the report said there was a “large consensus” internally that Gemini 4 is at the frontier.
Your reaction