tl;dr: Google released Gemini 4 Argon, a frontier model with 1M output tokens aimed at long-horizon scientific, coding, and agentic tasks, claiming top scores on benchmarks like Vals Index, Harvey's Legal Agent, LVBench, and tying for #1 on CWE-Bench for vulnerability remediation. Partners including Isomorphic Labs, Box, Deloitte, and Wiz report gains in drug discovery, enterprise analysis, and cybersecurity, and the model demonstrated sim-to-real transfer by teaching an Apptronik humanoid to play cornhole from video. It's available now to Google AI Pro and Ultra subscribers.