Skip to main content

Gemini 4 Argon Writes a Million Tokens in One Go

Google DeepMind's Gemini 4 Argon: A Million-Token Output and a Laser Focus on Complex Work

Google DeepMind quietly dropped Gemini 4 Argon on September 30, and the headline number is hard to ignore: a single output can now reach 1 million tokens. That's not the context window—it's how much the model can generate in one continuous stream. For anyone who's ever had to break a task into chunks, that's a game-changer.

Why Output Length Matters

Most cutting-edge models cap output at around 128,000 tokens. Argon's jump to a million means it can write tens of thousands of words—or migrate an entire codebase—without stopping to catch its breath. Deep codebase migrations, long technical reports, multi-step automation: these tasks suddenly gain a continuity that was previously impossible. No more stitching together segments and hoping they align.

Image

Built for the Hard Stuff

Argon isn't for casual chat. Google is rolling it out in phases, starting with trusted cybersecurity professionals before expanding to developers, enterprises, and subscribers. Its target? Complex, long-term workflows in software engineering, law, finance, and cybersecurity defense.

The benchmarks back up the ambition. DeepSWE v1.1, which tests long-term software engineering tasks, hit a record 77.9%. The Vals Index—weighted across finance, programming, law, and taxation—ranked Argon first. Zapier's AutomationBench gave it a 51.3% score, also first. And on long video understanding, LVBench, it reached 91.7%, the highest yet. Visual tasks are covered too: professional charts, extracting details from lengthy videos, and cross-document actions.

Security That Leans on Defense

Cybersecurity is the core narrative here. Argon is trained to autonomously discover, verify, and fix critical vulnerabilities. For trusted defenders and Google's internal teams, there's even a "no network barrier" version. Security vendor Wiz used it through the Scan for Good project to spot a serious flaw in a global hospital's medical software—one that previous cutting-edge models had missed. On CWE-bench v1, which tests vulnerability repair, Argon ranked first with 68%.

Pricing and Practicalities

During the introductory period, Argon costs $2 per million input tokens and $10 per million output tokens. After that, it doubles to $4 and $20. Before a wider release, Google is reinforcing four advanced defenses: preventing abuse, blocking prompt injection, monitoring misaligned reasoning chains, and strengthening sandboxes. Thousands of Google employees are already using it to debug code, design algorithms, and perform large-scale code migrations. In one case, it improved quantum algorithms by 40% in minutes.

So no, Argon isn't aimed at chatting. It's aimed at the hardest challenges professionals face—and it's swinging big.

Key Points

  • 1 million token output: Eight times the typical 128K limit, enabling uninterrupted long-form generation.
  • Targeted at professionals: Focused on software engineering, law, finance, and cybersecurity defense.
  • Top benchmark scores: 77.9% on DeepSWE v1.1, first on Vals Index, 51.3% on AutomationBench, 91.7% on LVBench.
  • Security-first: Trained to find and fix vulnerabilities; already caught a hospital software flaw.
  • Tiered pricing: $2/$10 per million tokens initially, rising to $4/$20 later.