Google announced Gemini 4 Argon on 30 September 2026, its first new frontier flagship since the Gemini 3 series in November 2025 — and it is deliberately holding the model back from general release.
The model goes first to "trusted cyber defenders" through Google's Fairwind Program, a limited-access track introduced earlier in September. Google says it is simultaneously taking part in the US government's voluntary pre-release model access process. Paid API customers and Google AI Ultra subscribers are next, with broad availability promised "as soon as possible," but no date given.
The benchmark picture is Google's strongest claim in months. Across the 18 benchmarks the company disclosed, Argon posts the top or joint-top score in 13 categories: 12 outright wins and one tie. GPT-6 Astra leads outright on three and ties Argon on one, while Claude Opus 5.5 leads two. Argon's widest margins come in enterprise and long-context work — 19.6% on Harvey's Legal Agent Benchmark against 5.4% for Astra and 3.8% for Opus, 51.3% on AutomationBench, 84.2% on GraphWalks and 91.7% on the long-video benchmark LVBench. It also shades the field on DeepSWE v1.1 (77.9%) and ties Astra at 68% on the vulnerability-remediation test CWE-bench v1.
It is not a clean sweep. Astra leads by 10.5 points on FrontierSWE v2 (65.5% to 55.0%) and on Terminal-Bench Science 0.1, while Claude Opus 5.5 is ahead on Terminal-bench 4.0 and PostTrainBench. The frontier race remains workload-dependent.
Argon also raises the output ceiling to 1 million tokens, up from 64,000 — a meaningful change for agentic work that runs long chains of steps. Google says the model is already used internally by thousands of employees, and cites a libgav1 job in which Argon agents replaced 32,000 lines of SIMD code with a memory-safe Rust decoder that runs 2.7 times faster. For cyber defenders Google is shipping Argon without the usual cyber guardrails so they can use its full defensive range; security firm Wiz says Argon uncovered a critical vulnerability in hospital software used worldwide.
Introductory API pricing is $2 per million input tokens and $10 per million output tokens, with cached input at $0.10, rising to $4/$20 later. That undercuts GPT-6 Astra at $10/$50 and matches Claude Opus 5.5's post-introductory rate.
The caveat: most customers cannot test any of it yet. Until broader API access opens, Argon's lead is a promise, not a product.
