Google DeepMind launched its new cutting-edge model, Gemini 4 Argon, on September 30. The most shocking change is the output limit: it has been increased from the original 64K to a staggering 1 million Tokens - note that this refers to "how long it can generate at once," not the context window. When the model can write tens of thousands or even millions of words in a single task chain without needing to break it into segments, deep codebase migration, long technical reports, and multi-step automation processes gain continuity from start to finish. Most cutting-edge models currently have an output limit of about 128,000 tokens, and Argon's jump is truly groundbreaking.
Enterprise-Level Long-Range Capabilities, Outstanding Evaluation Results
In terms of positioning, Argon follows a phased, targeted release approach, rather than being fully open from the beginning: first for trusted cybersecurity professionals, then expanding to developers, enterprises, and subscription users. Google positions it as focusing on "complex, long-term workflows," targeting software engineering, law, finance, and cybersecurity defense. The results are impressive: DeepSWE v1.1 achieved a record 77.9% in long-term software engineering tasks; Vals Index (weighted by finance, programming, law, and taxation) ranks first; Zapier's AutomationBench scored 51.3% to rank first; and long video understanding LVBench reached 91.7%, the highest. On the visual side, it can also handle professional charts, extract details from long videos, and perform cross-document actions.
Security Focus on Defense, Pricing at Both Ends
Cybersecurity is the core narrative of this release. Argon is trained to autonomously discover, verify, and fix critical vulnerabilities. For trusted defenders and Google's internal teams, it will also provide a "no network barrier" version. Security vendor Wiz used it through the Scan for Good project to identify a serious vulnerability in a global hospital medical software, which even previous cutting-edge models had failed to detect. CWE-bench v1 vulnerability repair ranked first with 68%.
The pricing has two tiers: during the introductory period, it costs $2 per million input tokens and $10 per million output tokens, and afterward, it will increase to $4 and $20, respectively. Before making it widely available, Google is reinforcing four advanced defenses — preventing abuse, preventing 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. Quantum algorithms were improved by 40% in minutes with its help. In short, Argon is not aimed at chatting, but at tackling the hardest challenges for professionals.
