Google Launches Gemini 4 Argon to Restricted Testers

Google DeepMind announced Gemini 4 Argon on September 30, 2026, making it available first through the Fairwind Program to trusted testers, internal staff, and vetted cybersecurity defenders. API customers and AI Ultra subscribers are next in line before any broader public access. Google is also engaged in the U.S. government's voluntary pre-release model access process as part of a phased rollout.

Google claims Argon scores 77.9% on DeepSWE v1.1, 91.7% on LVBench, 51.3% on AutomationBench, and 68% on CWE-bench v1. Across 19 benchmarks, it leads outright on 13, but GPT-6 Astra leads on several software, science, and computer use benchmarks, and Claude Opus leads Terminal-Bench and PostTrainBench. On finance and legal agent benchmarks, Argon scores 65.4% versus GPT-6 Astra's 53.5% on finance, and 19.6% versus 5.4% on legal. Vals AI, using its own index, ranks Argon first out of 41 models at 68.90% accuracy. No outside lab has independently replicated Google's benchmark figures.

The model's output token ceiling is 1 million tokens, which one report describes as nearly 8x the 128K cap on competing models including GPT-6 Astra, Claude Opus 5.5, and Fable 5.1. On prompt injection resistance, Google reports a 0.7% attack success rate after 15 attempts on the Gray Swan IPI benchmark, compared to 52.7% for Kimi K3.

Internally, Google reports Argon agents freed over 300 TiB of memory across its data centers, with total estimated savings of 500 TiB to 1 PiB. Argon is also migrating C/C++ codebases to Rust, including over 800,000 lines in the Fuchsia OS Zircon kernel and work on the re2 and libgav1 libraries. On libgav1, Argon replaced 32,000 lines of SIMD code, making the Rust port 2.7x faster.

Pricing is $2 per million input tokens and $10 per million output tokens, with cached input at 95% off. For cybersecurity applications, Google says trusted defenders and internal teams will receive Argon without cyber guardrails to use its full capabilities, with chain-of-thought inspection and hardened sandboxed environments as safety measures. Google had previously promised a Gemini 3.5 Pro release in June 2026 but instead released smaller Flash models over the summer; the last of those, Gemini 3.8 Flash, shipped September 2, 2026.