Google DeepMind launched the next-generation cutting-edge model Gemini 4 Argon on September 30. The most significant change is the upper limit of output: it has been increased from the original 64K to 1 million Tokens—note that this refers to how long a single generation can be, not the context window. When the model can write hundreds of thousands or even millions of words in a single task chain without分段 requests, the deep code library migration, long technical reports, and multi-step automated processes become coherent from start to finish. Most cutting-edge models currently have an output limit of approximately 128,000 Tokens, and the Argon hop can be considered a major breakthrough.
Enterprise-level long-distance tasks, leading in comprehensive evaluation
In terms of positioning, Argon adopts a phased and targeted release approach, rather than making it available to everyone from the start: first for trusted network defenders, then for developers, enterprises, and subscription users. Google has placed it within the "complex, long-term workflow" framework, focusing on software engineering, law, finance, and network security defense. The report is quite impressive: DeepSWE v1.1 long-term software engineering achieved 77.9% to set a record; the Vals index (weighted for finance / programming / law / taxation) topped the list; Zapier’s AutomationBench ranked first with 51.3%; and long video comprehension on LVBench reached 91.7%, the highest. From a visual perspective, it also allows processing professional charts, extracting details from long videos, and performing actions across documents.
The security card focuses on defense, with pricing set at both ends
Network security is the core narrative of this release. Argon has been trained to autonomously detect, verify, and fix critical vulnerabilities; for trusted defenders and Google internal teams, a “no-network barrier” version will also be released. Security vendor Wiz, through the Scan for Good project, identified a serious vulnerability in a global hospital medical software that no advanced model had previously detected. The vulnerability fix using CWE-bench v1 ranked first with 68%.
The pricing is divided into two tiers: 2 dollars per million inputs and 10 dollars for outputs, which later increase to 4 dollars and 20 dollars. Before becoming widely available, Google also strengthened four frontline defenses—against abuse, against prompt injection, monitoring of misaligned thought chains, and the enhancement of the sandbox. Thousands of Google employees use it to debug code, design algorithms, and perform large-scale code migration; quantum algorithms can improve benchmarks by 40% in just a few minutes through it. In short, Argon is not aimed at chatting, but at helping professionals tackle the most difficult challenges.