The lede
Gemini 4 Argon: our next era of frontier intelligence
Google DeepMind
Google's new frontier model is going first to trusted cyber defenders through a gated program rather than to developers, which is a story in itself, but the numbers to plan around are the output ceiling moving from 64K to 1M tokens, $2 and $10 per million in and out with cached input at 95% off, and a new state of the art on DeepSWE v1.1 at 77.9%. The internal receipts are more persuasive than the benchmarks: Argon agents took an existing Rust port of libgav1, replaced 32K lines of hand-written SIMD with safe Rust the compiler vectorizes on its own, and came out 2.7x faster than that port with identical video output, while another fleet of agents freed over 300 TiB of memory across Google's data centers from fleet-wide profiling telemetry. The interesting claim is not intelligence, it is headroom — a model handed a million output tokens in a single trajectory is being asked to finish things rather than answer them.