Google announced Gemini 4 Argon on September 30, 2026 — a frontier model built for sustained deep reasoning across long-horizon workflows. The headline capability is a 1M output token limit (up from 64K), enabling the model to think through complex multi-step problems in a single trajectory. Argon is already deployed internally across Google, where thousands of engineers use it for coding, research, and large-scale codebase migrations. The numbers Google chose to highlight are pointed. Argon agents freed 300+ TiB of memory across Google's data centers, with estimates of 500 TiB to 1 PiB total. They're migrating C/C++ codebases to Rust at scale — up to 800K+ lines for the Fuchsia Zircon kernel — and in one case produced a memory-safe video decoder that runs 2.7x faster than the existing Rust port. On quantum computing, the model beat a published baseline by 40% in minutes. These aren't benchmark scores; they're internal productivity claims that double as competitive moats. On external benchmarks, Argon posts 77.9% on DeepSWE v1.1 (real-world software engineering), leads the Vals Index across finance/legal/tax weighted by GDP contribution, tops Zapier's AutomationBench at 51.3%, scores 91.7% on LVBench (long video understanding), and ties for first on CWE-bench v1 (vulnerability remediation) at 68%. Google is claiming state of the art across coding, enterprise knowledge work, and cybersecurity defense simultaneously. The rollout strategy is the real story. Argon ships first to a curated set of cyber defenders through Google's Fairwind Program, with guardrails explicitly removed for that cohort so they can leverage full offensive-grade capabilities. Google is participating in the U.S. government's voluntary pre-release model access process. Broader developer, enterprise, and consumer access comes later — no date given. This phased approach lets Google control the narrative on safety while establishing first-mover relationships with high-value institutional buyers. Pricing is set at $2 per million input tokens and $10 per million output tokens, with cached inputs at 95% off. This is aggressive — designed to undercut competitors and establish Argon as the default API for enterprise workflows. At these rates, the model is priced to acquire market share, not to maximize per-token revenue. The 95% cache discount in particular rewards sticky, high-volume enterprise integrations. The safety section is detailed but structured to reassure, not to constrain. Google describes CBRN refusal guardrails, prompt injection resistance (leading on Gray Swan's IPI benchmark), chain-of-thought monitoring for misalignment, and hardened sandboxed environments. Notably, they call on the industry to 'preserve reasoning transparency' — a strategic move that simultaneously signals responsibility and pressures competitors who might consider hiding chain-of-thought. The monitoring system was explicitly firewalled from training to avoid teaching the model to evade detection. The cybersecurity vertical deserves separate attention. Wiz is already using Argon through its Scan for Good initiative, where the model uncovered a critical vulnerability in healthcare software exposing patient data across hospitals worldwide — a finding previous frontier models missed. Google is positioning Argon as a national security asset, which creates a procurement pathway that bypasses normal enterprise sales cycles and locks in government and critical infrastructure relationships before competitors can respond.
