
Google introduces Gemini 4 Argon with million-token output
Google has announced Gemini 4 Argon, its new frontier AI model for long running tasks in software development, enterprise work, and cybersecurity. Its maximum output rises from 64,000 to 1 million tokens, giving it more room to handle longer reasoning and complex workflows in a single run. Google reports scores of 77.9% on DeepSWE v1.1, 51.3% on Zapier's AutomationBench, and 91.7% on LVBench, and says Argon leads the Vals Index while outperforming GPT-6 Astra and Claude Opus 5.5 on several evaluations.
Access is currently limited to selected cybersecurity defenders through Google's Fairwind Program while safety testing continues. Google says Argon can discover vulnerabilities, validate exploitability, and generate patches, with its Wiz team already using the model through Scan for Good. Argon also scored 68% on CWE-bench v1, tying for first place.
Access will later expand to paid API customers and Google AI Ultra subscribers. Introductory pricing is $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached inputs, before increasing to $4 and $20 respectively. Google has not announced a date for broader availability.






Comments
I’m really looking forward to trying this. A million-token output limit is impressive, but what interests me more is whether the model can stay coherent, follow requirements, and verify its work throughout a long task.
For software development and automation, the real gain would be spending less time correcting an agent or helping it recover context between steps. If Argon delivers on that, it could make a meaningful difference to how I build projects. Eager to see how it performs on my own workflows.
It's getting hard to judge which LLM subscription will be the best for your individual use case. This Gemini 4 costs more than GPT-6.1-Sol for solving tasks, but scores much higher in the APEX-Agents benchmark.
For me it comes down to personal preference: don't give Google, Microsoft, or Meta any data if I can help it. So I stay with Claude/OpenAI.