Google's Gemini 4 Argon arrives with 1M-token output and a cyber-defenders-first rollout

Google announced Gemini 4 Argon on Wednesday, September 30, 2026, its first new frontier model in more than seven months and the long-awaited follow-up to the Gemini 3 generation. In a post on Google's Keyword blog, Koray Kavukcuoglu, SVP of Google DeepMind and Google's Chief AI Architect, described Argon as a model for "complex workflows across real-world software engineering, enterprise knowledge work like legal and finance, and cybersecurity defense." There is a catch: almost nobody can use it yet. Argon is rolling out first to a group of "trusted cyber defenders" through a new Fairwind Program, and Google has given no date for broad availability.

Axios called it Google's "long-awaited answer to OpenAI and Anthropic," noting that the company spent much of the past year shipping smaller, cheaper Flash models while its rivals pushed forward at the top end. The Decoder reports that Argon follows Gemini 3.1 Pro as Google's last frontier release, and that an already-announced Gemini 3.5 frontier model was skipped entirely during a drawn-out development period.

The headline feature: one million output tokens

The most concrete technical change is the output limit. Per Google, Argon can generate up to 1 million tokens in a single response, up from 64,000 on previous Gemini models. Google calls this industry-leading, and its pitch is that a model with room to "generate hundreds of thousands of tokens in a single trajectory" can reason through hard, long-horizon problems in one go instead of being chopped into many shorter calls.

According to The Decoder, Google is adding a "Long Decode Continuation" feature to the Gemini API to make that practical: it pauses very long responses and resumes them through follow-up requests so they don't hit a timeout. The input context window stays at one million tokens, and Argon accepts text, images, video and audio as input but outputs only text.

What Google says it beats

Google's own numbers are strong. The company says Argon sets a new state of the art on DeepSWE v1.1, a benchmark of long-horizon real-world software engineering tasks, at 77.9%. It says Argon is the top model on the Vals Index, which weights finance, coding, legal and tax work by each sector's contribution to U.S. GDP; Vals AI puts its score at 68.9%, per The Decoder, making it the first Gemini model to top that index. Google also claims first place on Zapier's AutomationBench (51.3%) and on the LVBench long-video understanding test (91.7%).

VentureBeat, summarizing the comparison table Google published, reports that Argon leads or ties on 13 of the 18 benchmarks Google disclosed against OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 — outright first on 12 and tied on one. That same table isn't a clean sweep, though. Claude Opus 5.5 still wins Terminal-bench 4.0, and GPT-6 Astra leads on FrontierSWE v2.

What independent testing says

The early third-party picture is more mixed. Per The Decoder, Artificial Analysis scores Argon at 53 on its Intelligence Index at the model's highest available reasoning setting — tied with GPT-6 Astra and Claude Fable 5.1, but behind Claude Opus 5.5 (58) and Claude Sonnet 5.5 (56). That is still a 23-point jump over Gemini 3.1 Pro Preview.

Two details from that testing are worth flagging. First, the low price per token doesn't mean low cost per job: Artificial Analysis found Argon uses an average of about 62,000 output tokens per task, versus roughly 27,000 for GPT-6 Astra. At the introductory price that still works out cheaper per task than Astra, but once pricing doubles, The Decoder reports Argon becomes about 20% more expensive than Astra on the same workload. Second, Argon posted a notably low hallucination rate of 15% on Artificial Analysis's AA-Omniscience test, against 51% for GPT-6 Astra — it is much more willing to say it doesn't know — though its raw accuracy on that test was lower.

On agentic tasks, long a weak spot for Gemini, Artificial Analysis measured Argon at 57% on Terminal Bench 4, a 53-point jump over Gemini 3.1 Pro Preview but still behind Claude Sonnet 5.5, Claude Opus 5.5 and GPT-6 Astra, according to The Decoder. Human-preference rankings were kinder: on Arena.ai's Text Arena, Argon took first place, though it landed eighth in the WebDev coding arena. The Decoder's summary headline captures it: Argon "closes the gap with OpenAI and Anthropic but doesn't take a clear lead."

Why cyber defenders get it first

The rollout is the most unusual part of the launch. Google says Argon was trained to be highly capable at cybersecurity defense — able to "autonomously find, validate, and patch critical software vulnerabilities" — and that trusted defenders in the Fairwind Program and Google's own internal teams will get the model without cyber guardrails. Google also says it is taking part in the U.S. government's voluntary process for pre-release model access.

"Starting this rollout in this way gives us more confidence, but also enables us to put a model that is trained and strong in cyber defense in the hands of defenders as soon as possible," Tulsee Doshi, Google's Gemini model product lead, told CNBC. CNBC also noted the launch came a day after CEO Sundar Pichai signed a voluntary AI safety agreement with President Donald Trump and other tech executives at the White House.

Google cites one early real-world result: security firm Wiz, using Argon through its Scan for Good program, found a critical vulnerability exposing sensitive personal information in healthcare software used by hospitals worldwide, a flaw Google says previous frontier models had missed. On CWE-bench v1, which measures vulnerability remediation, Google says Argon ties for first at 68%.

Before a wider release, Google says it is hardening safeguards in four areas: blocking misuse for cyber and CBRN attacks, resisting indirect prompt injection, monitoring the model's chain-of-thought and actions for misalignment, and sealing the sandboxed environments used for high-risk training and evaluation. General access will come later, starting with paid API customers and Google AI Ultra subscribers; Google says only "as soon as possible."

What Google says Argon is already doing internally

Google's announcement leans heavily on internal use. It says a team of Argon agents analyzed fleet-wide profiling data and applied memory optimizations across Google's data centers, freeing more than 300 TiB of memory once rolled out, with an estimated 500 TiB to 1 PiB in total savings. Argon agents are also migrating C and C++ codebases to Rust, from core libraries like re2 and libgav1 up to the 800,000-plus-line Fuchsia Zircon kernel. For libgav1, Google's open-source video decoder, Google says Argon replaced 32,000 lines of SIMD code with safe Rust the compiler could vectorize on its own, producing a decoder 2.7x faster than the existing Rust port with identical output. These are Google's own claims and haven't been independently verified.

What it means

For developers, the practical news is mostly the price card and the 1M-token output ceiling — neither of which you can use today. For the industry, Argon matters because it puts Google back in the frontier race after a long stretch of Flash-only releases. Google's benchmarks say it leads; the first independent numbers say it is competitive with GPT-6 Astra and still behind Anthropic's top Claude models on overall intelligence scores. The real test is what happens once paying customers can run it on their own workloads — and how long the half-price introductory period lasts when they do.