Google DeepMind Unveils Gemini 4 Argon with 1M Output Tokens
Google DeepMind introduces Gemini 4 Argon for coding, knowledge work, and cybersecurity, featuring 1M output tokens.
Stock photo for illustration only, not from the actual event
- Google DeepMind launches Gemini 4 Argon supporting up to 1 million output tokens
- Introductory pricing starts at $2 per 1M input tokens and $10 per 1M output tokens
- Outperforms rivals on 12 of 18 benchmarks and patches critical software flaws
- Wiz utilizes the model via Scan for Good to discover a critical healthcare software vulnerability
Google DeepMind has officially introduced its newest artificial intelligence model, Gemini 4 Argon, engineered specifically to handle complex enterprise workflows ranging from software development and knowledge tasks to cybersecurity defense. The model is designed to deliver deep reasoning capabilities over extended sequences.
Adopting a phased deployment strategy, the company is participating in the United States government's voluntary framework for pre-release model access. This approach allows the development team to gather early tester feedback and refine safety guardrails prior to a broader public rollout.
Regarding commercial pricing, Argon launches with an introductory rate of $2 per million input tokens and $10 per million output tokens. Cached input tokens benefit from a 95% discount, reducing the cost to $0.10 per million. Logan Kilpatrick confirmed these introductory figures before prices transition to $4 for inputs and $20 for outputs.
Stock photo for illustration only, not from the actual event
Current industry frontier APIs cap individual model responses much lower, with models like Claude Opus 5.5, Claude Fable 5.1, and GPT-6 Astra restricted to 128K output tokens. According to the Google team, Argon can perform deep reasoning and generate hundreds of thousands of tokens within a single trajectory, allowing developers to execute massive code refactoring or draft exhaustive reports without fragmenting tasks across turns.
"Argon can think deeply and generate hundreds of thousands of tokens in one trajectory."
Google team
When evaluated against GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1, Argon claims outright leadership across 12 out of 18 evaluation benchmarks while tying for first place in one. Furthermore, Artificial Analysis reported that Argon matches GPT-6 Astra's Intelligence Index while operating at only 60% of the cost per task using discounted pricing tiers.
Expanding output capacity to 1 million tokens represents a major architectural milestone, enabling continuous end-to-end processing of extensive codebases and complex documents. Nonetheless, the absolute financial cost of utilizing maximum output lengths remains a critical variable for enterprise budget planning, alongside the imperative safety controls needed for sensitive cybersecurity applications.
Additionally, Argon has been trained to autonomously discover, validate, and patch critical software security flaws. Trusted defenders and internal Google teams have been granted access without standard cyber guardrails. On the CWE-bench v1 vulnerability remediation benchmark, Argon ties for the top position at 68%.
Wiz has already integrated Argon through its Scan for Good initiative, successfully identifying a severe vulnerability in healthcare software utilized by hospitals globally that previous frontier models failed to detect. Before a general rollout, Google is actively hardening safeguards across four key operational areas, backed by usage data from thousands of internal employees.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment