Gemini 4 Argon Beats GPT-6 Astra at 40% the Cost

Google has introduced Gemini 4 Argon, its most advanced AI model to date, specifically designed for coding, research, enterprise applications, and cybersecurity tasks. Despite matching the performance of GPT-6 Astra in various benchmarks, Argon is significantly more cost-effective, costing 40% less than its counterpart. Currently, Google is offering Argon exclusively to a select group of cyber defenders via its Fairwind Program before gradually opening up access to developers, enterprises, and consumers.

Pricing for Argon starts at $2 per million input tokens and $10 per million output tokens during the initial rollout period, which is expected to increase to $4 and $20 respectively after the introductory period. Built for intricate, long-term workflows, Argon is already being utilized internally at Google for debugging, research, algorithm design, and large-scale code migrations.

The model achieved a remarkable 77.9% score on the DeepSWE v1.1 benchmark, which focuses on real-world software engineering tasks, as well as a top score of 51.3% on AutomationBench, where it outperformed other models. Additionally, Argon demonstrated exceptional performance on the LVBench, which assesses long-video understanding, with a score of 91.7%.

It successfully assisted in migrating large C/C++ codebases to Rust, including projects with over 800,000 lines of code, and optimized a Rust video decoder, increasing its speed by 2.7 times compared to a previous Rust implementation. Independent evaluations rank Gemini 4 Argon as the top model out of 41 on the Vals Index, scoring 68.90%, surpassing Claude Sonnet 5.5 (67.04%), Claude Opus 5.5 (66.97%), and GPT-6 Astra (63.13).

However, it does not outperform all benchmarks, as it trails GPT-6 Astra in some individual tests. Trained specifically for defensive cybersecurity, Argon possesses the capability to autonomously identify, validate, and rectify software vulnerabilities. In a recent incident, Wiz utilized the model to detect a critical vulnerability in healthcare software that had previously gone undetected by other frontier models.