Google Unveils Gemini 4 Argon with 1 Million Token Output

Google DeepMind has released Gemini 4 Argon, a powerful new AI model designed for heavy-duty work like coding, finance, law, science and cybersecurity ¹. The biggest news is its output limit: Argon can write up to one million tokens in a single response, a massive jump from the 64,000-token cap on earlier Gemini versions.

How It Works

Argon is built on a sparse mixture-of-experts design, a type of architecture that only activates the parts of the model needed for each task rather than firing up the whole system at once. This should make it faster and cheaper than models that use every parameter all the time. The model accepts text, images, audio and video, but only outputs text. Google trained it on its own TPUs using its JAX framework and ML Pathways system.

The one-million-token limit covers both input and output combined, so a very long prompt will eat into the space left for the model's response. A full one-million-token reply costs $10 in output tokens at intro pricing, along with $2 per million input tokens. After November 30, 2026, those prices double to $4 and $20.

Benchmarks: Strong but Not Everywhere

Google says Argon leads on several key benchmarks. It scored 68.90 percent on the Vals Index, ranked first by Vals AI. It also topped DeepSWE v1.1 at 77.9 percent and AutomationBench at 51.3 percent, according to Google's own comparisons ¹.

But the model stumbles on some real-world coding tests. It scored just 55.0 percent on FrontierSWE v2, trailing GPT-6 Astra at 65.5 percent, and ranked last among the four models in Google's comparison table. It also underperformed on Terminal-Bench 4.0 and OSWorld-2.0. Google acknowledges that Argon can occasionally produce hallucinations and experience slowdowns or timeouts.

Pricing and Access

Introductory pricing makes Argon cheaper than both Claude Opus 5.5 and GPT-6 Astra for now. Cached input tokens come with a 95 percent discount. The model is available through the paid Gemini API, AI Studio, Android Studio, the Gemini platform for enterprises, and Google Antigravity. Consumers with AI Pro or Ultra subscriptions can also access it through the Gemini app and AI Mode in Search. Cyber defenders can try it through the Fairwind Program without standard cyber guardrails.

What Google Already Built It For

Before the public launch, Google used Argon inside its Antigravity coding tool, claiming it boosted agentic code submissions by over 35 percent and cut rollbacks by nearly a third ¹. In its open-source Agent Development Kit, the model reduced session preparation time by 40 percent and made cold imports 21 percent faster. Google also pointed to a dramatic drop in successful indirect prompt injection attacks on its Gray Swan benchmark, falling to just 0.7 percent across 15 adaptive attempts.

Gemini 4 Argon joins a crowded frontier model market where speed, output length and raw capability are now all racing upward. Whether it stays on top will depend on whether developers find it reliable enough for the kinds of long, multi-step projects it was built for.

Key facts

Gemini 4 Argon can output up to 1 million tokens in a single response
Sparse mixture-of-experts architecture activates only needed model parts
Intro pricing is $2 input / $10 output per million tokens until November 30, 2026
Trails GPT-6 Astra on FrontierSWE v2 coding benchmark at 55.0 percent
Ranked #1 on Vals Index at 68.90 percent