Originally published on The Daily Flare.
Google has launched Gemini 4 Argon, the first model in its Gemini 4 line, with a one-million-token output limit and early access reserved for vetted cyber defenders.
After weeks of release date speculation, Google announced Gemini 4 Argon on September 30, 2026, positioning it as the company's answer to OpenAI's GPT-6 family in the intensifying AI development race. Built by Google DeepMind, the model is designed for long-horizon work: real-world software engineering, enterprise knowledge tasks spanning legal and financial research, and autonomous cybersecurity defense. Google describes it as the start of a new era of frontier intelligence.
A million-token output window
Argon's headline feature is its output limit of one million tokens, up from 64,000 in the previous generation. That allows the model to produce hundreds of thousands of tokens in a single pass — an entire codebase refactored, a long legal brief drafted, or an extended vulnerability analysis written without stopping for breath.
Google says the capability rests on a serving feature called Long Decode Continuation, which pauses very long responses and resumes them across later calls so requests do not time out. The model accepts text, images, and video as input with up to a million tokens of context, and offers adjustable reasoning effort so users can trade speed for depth.
Gemini 4 Argon benchmarks against GPT-6 Astra and Claude
On Google's own benchmark table, Argon leads or ties 13 of 19 rows against OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 and Fable 5.1. Standouts include 77.9% on DeepSWE v1.1, a test of real-world long-horizon software engineering, 91.7% on LVBench for long video understanding, and 68% on the CWE-bench security benchmark, where it ties for first place.
Independent checks broadly agree the model is frontier-class. Artificial Analysis scores Argon 53 on its Intelligence Index, level with GPT-6 Astra, and security firm Wiz reports the model found a critical flaw in hospital software that earlier frontier models had missed.
There are trade-offs. Independent analysis finds Argon spends roughly twice as many tokens per task as its closest rival, which makes it costlier and slower per completed task despite identical per-token pricing — around three times the wall-clock time in agent loops where every turn waits on the model.
Pricing and who gets access first
Argon launches at an introductory price of $2 per million input tokens and $10 per million output tokens, doubling to $4 and $20 when the introductory period ends. Google has not said when that happens, and the model still has no public entry on the company's API pricing page.
Most people cannot use it yet. Google is releasing Argon first to vetted cyber defenders through its Fairwind Program, which it says counts over 650 partners, alongside its own engineers and the U.S. government under a voluntary pre-release process. Those early users receive a version without cybersecurity guardrails; the wider release will refuse harmful requests related to cyber, chemical, biological, radiological, or nuclear attacks.
Paid API customers and Google AI Ultra subscribers are next in line, followed by developers, enterprises, and consumers — on a date Google has not named, beyond "as soon as possible." The staggered rollout reflects the model's dual-use nature: the same capability that finds vulnerabilities in hospital software could be turned against critical infrastructure, which is why defenders get it first.
Under the hood, Argon runs on Google's TPU v6 Trillium clusters — warehouse-scale data centers built for exactly this kind of long-horizon computation. As with every new generation of large language models, the real test will be what developers build once the gates finally open.
Top comments (0)