Gemini 4 Argon: Google's New Model Writes a Million Tokens in One Go

What if an AI could write an entire codebase migration plan without ever stopping to catch its breath? That's the promise of Google DeepMind's latest model, Gemini 4 Argon, unveiled on September 30. Its headline feature: a staggering 1 million token output limit—up from the previous 64K. To put that in perspective, most cutting-edge models today cap out around 128,000 tokens. Argon doesn't just double that; it leaps eight times beyond.

And no, this isn't about the context window. This is about how much the model can generate in a single, uninterrupted output. That means deep codebase migrations, lengthy technical reports, and multi-step automation can now flow from start to finish without awkward breaks.

Built for the Long Haul

Google isn't positioning Argon as a chatbot for casual banter. Instead, it's aimed squarely at "complex, long-term workflows" in software engineering, law, finance, and cybersecurity defense. The rollout is phased: first to trusted cybersecurity professionals, then developers, enterprises, and eventually subscribers.

The benchmarks back up the hype. On DeepSWE v1.1, which tests long-horizon software engineering tasks, Argon scored a record 77.9%. It also topped the Vals Index (a weighted measure across finance, programming, law, and taxation) and Zapier's AutomationBench with 51.3%. For long video understanding, it hit 91.7% on LVBench—the highest yet. And visually, it can handle professional charts, extract details from lengthy videos, and perform cross-document actions.

Security That Goes on the Offense

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. In one case, security vendor Wiz used Argon through the Scan for Good project to uncover a serious flaw in a global hospital's medical software—a vulnerability that had slipped past previous cutting-edge models. On the CWE-bench v1 vulnerability repair benchmark, Argon ranked first with 68%.

Pricing comes in two tiers: during the introductory period, it's $2 per million input tokens and $10 per million output tokens; afterward, those rates double to $4 and $20. Before wide release, Google is beefing up four advanced defenses—preventing abuse, blocking prompt injection, monitoring misaligned reasoning chains, and strengthening sandboxes.

Already at Work Inside Google

Thousands of Google employees are already using Argon to debug code, design algorithms, and perform large-scale code migrations. In one striking example, quantum algorithms were improved by 40% in minutes with its help. Clearly, Argon isn't here to chat—it's here to tackle the hardest challenges professionals face.

Key Points

• 1 million token output: Argon can generate up to 1 million tokens in a single output, eight times the typical 128K limit.
• Targeted at professionals: Focused on complex, long-term workflows in software engineering, law, finance, and cybersecurity.
• Top benchmark scores: 77.9% on DeepSWE v1.1, first on Vals Index and AutomationBench, 91.7% on LVBench.
• Security focus: Autonomously finds and fixes vulnerabilities; already helped identify a hospital software flaw.
• Phased rollout: First to trusted cybersecurity pros, then developers and enterprises.
• Pricing: Introductory $2/$10 per million input/output tokens, later $4/$20.