Google releases Gemini 4 Argon with strong knowledge work performance

Google has released Gemini 4 Argon, a new frontier model built on a fresh pre-training run after the company shelved Gemini 3.5 Pro. Benchmarks indicate that Argon is state-of-the-art for knowledge work tasks, leading the VALS Index, Automation Bench, and Harvey's Legal Agent benchmark by 13 points over Fable 5.1. However, its performance in coding is mixed; while it leads on DeepSwee, it trails Astra by 10 points on FrontierSwee and Opus 5.5 by 9 points on Terminal Bench 4.0. On the OSworld benchmark, Argon scored 69.2%, slightly behind Astra's 72.6%. Artificial Analysis ranked Argon fourth with a score of 53, behind Opus 5.5 at 58 and tied with Fable 5.1 and GPT-6.

API pricing for the model includes a 50% discount until November 30, costing $199 per task on the Artificial Analysis benchmark. Performance varies significantly depending on the harness used; Pavel Huren noted that Argon performed near Opus 5.5 Max in the Antigravity CLI but poorly in the Gemini CLI. Opinions on the model's utility are divided, with Noah Dunnigan calling it useless for real-world code, while Dan Shipper praised its writing capabilities, noting that DeepMind hired professional writers to reduce AI slop. The model is also noted for unique multimodal and 3D generation capabilities that distinguish it from rivals.