Google launches Gemini 4 Argon: $1.99 per task, 1M-token output to reclaim the lead

新智元 · wechat · 2026-10-01

Per Xinzhiyuan, Google released Gemini 4 Argon, its first non-Flash proprietary model in 7 months. It tops TextArena (1533 score) and ties GPT-6 Astra on Artificial Analysis' intelligence index (53), while pricing at $2/M input and $10/M output tokens — a blended $1.99 per task, 60% of GPT-6 Astra's cost, with 95%-off cached input, widely read as a TPU-fueled price war.

Output limit is raised to 1M tokens. Internal deployments reportedly freed 300+ TiB of memory, rewrote 800K lines of C/C++ kernel code into Rust, and rewrote 32K lines of SIMD code in libgav1 for a 2.7x decode speedup. On security, Argon autonomously found a critical vulnerability in widely used medical software; Google is using tiered release — a no-guardrail version only for trusted defenders via the Fairwind program, restricted API and public access later. The report also covers the scrapped Gemini 3.5 Pro ($400M training cost), researcher departures, and a three-way race with OpenAI and Anthropic.

Original post →

Google launches Gemini 4 Argon: $1.99 per task, 1M-token output to reclaim the lead

More from Models