Google's Gemini 4 Argon lands in limited release — and retakes the frontier benchmarks lead

Google DeepMind has announced Gemini 4 Argon, its new frontier model, on Sep 30, 2026 — and it is doing something unusual for a launch this big: it is initially gated. Argon is rolling out first to a cohort of trusted cyber defenders through Google's Fairwind Program while the company engages in the U.S.’s voluntary pre-release access process for frontier models, with developers, enterprises and consumers scheduled for broad availability “as soon as possible.”

Limited release, half-price intro

The pricing is the most notable detail. Argon launches at an introductory $2 per million input tokens and $10 per million output tokens, with cached input priced 95% off. That introduces a 50% discount; once it expires, Google's published rate becomes $4 / $20 per million tokens. That landing price matches the exact rate Meta attached to its Muse Wonder frontier model this month — two labs converging on a similar per-million-token bill at the top of the stack.

One million output tokens, not input

The headline capability is headroom for long-horizon work: Argon expands its output limit to an industry-leading 1M tokens, a big leap from the previous 64K. A single trajectory can now run hundreds of thousands of tokens without the model dropping the thread — exactly what sustained multi-step reasoning on a complex engineering task or a long research brief requires. Artificial Analysis notes it has arrived as Google's first proprietary model above the Flash class in over seven months.

Where it leads on numbers

Google's own published benchmarks put Argon at the state of the art on DeepSWE v1.1 (77.9%) for real-world long-horizon software engineering, and leading on domain-specific evaluations including the Vals Index (economic impact across finance, coding, legal and tax), AutomationBench (51.3%) and LVBench long-video understanding (91.7%).

Artificial Analysis's independent Intelligence Index corroborates the frontier claim: with high reasoning enabled, Argon scores 53 — matching GPT-6 Astra and edging out GPT-6.1 Sol (52) — while at its discounted rate it costs about $1.99 per index task, roughly 60% of Astra's cost.

Cyber angle

Proof in the field

Wiz is already running Argon through its Scan for Good initiative, where the model surfaced a critical vulnerability exposing sensitive personal information across hospital healthcare software worldwide — a severe risk earlier frontier models missed. Google also describes internal wins: a quantum-optimization example beating the published baseline by 40% in minutes, and fleet-wide memory optimizations freeing over 300 TiB (with an estimated 500–1,000 TiB total).

What this means for you

Security teams: if you are building security tooling, Argon's ungated release through Fairwind is a rare early-access window on a frontier model explicitly aimed at vulnerability discovery and patching — worth engaging now before broad availability. Engineers and enterprises: the 1M-output headroom plus top DeepSWE score makes it a serious candidate for long-running agentic coding and codebase migration, but the gated rollout means production integration should wait until Google confirms public API timing.

Watch next: the end date of the 50% introductory discount, when Argon opens to paid API customers and consumers, and whether the benchmark lead holds in independent, real-world evaluation once it leaves the Fairwind cohort.

Sources: Google — Introducing Gemini 4 Argon (Sep 30, 2026); VentureBeat — Google unveils Gemini 4 Argon, retaking benchmark lead in limited release (Sep 30, 2026); Artificial Analysis — Gemini 4 Argon: Google is back as one of the top three labs (Sep 30, 2026).

Corrections: Report an error at corrections@wireshaman.com.