IT News on October 1st: Google released Gemini 4 Argon yesterday (September 30th), claiming it to be the company's most advanced AI model to date, but it is not yet fully available to the public.
In terms of positioning, Gemini and Argon are Google's current "most advanced" models, focusing on long-duration software engineering, corporate knowledge management, and network security defense. Officials claim that they have set records in real-world software engineering testing, tied for first place in network security benchmarks, and lead in benchmarks for professional tasks such as finance and law.
In terms of model characteristics, Gemini Argon can output up to approximately 1 million tokens at a time, which is higher than the previous limit of 64,000 tokens. This makes it suitable for handling large codebases, long documents, and multi-stage tasks.
Google's DeepMind Gemini product leader, Tulsee Doshi, stated: "The Argon is a model with comprehensive capabilities and possesses cutting-edge abilities in multiple fields."
The new model is at the forefront in certain coding and knowledge work benchmark tests, outperforming its competitors OpenAI in multiple aspects, including the leading models GPT-6 and Astra.
In terms of performance, Google claims that its DeepSWE v1.1 has a score of 77.9%. The version with a score of 5.5, Claude Opus, has a score of 74.2%, and GPT-6 Astra also has a score of 74.1%.
In terms of security defense performance, Google trained Gemini and Argon to make them 'outstanding in network security defense', which represents a significant improvement over the 3.8 version of Flash Cyber. In the CWE-bench and v1 tests that evaluate the models' ability to fix security vulnerabilities, Argon tied for first place with the highest score of 68%.
In terms of distribution, Google stated that it will release the product in phases. In the first phase, it will be provided to trusted cybersecurity defenders through a program named Fairwind Program, and they will also participate in the voluntary pre-release model access process with the U.S. government.
In terms of fees, Google plans to initially offer it to paid API customers and Google AI Ultra users, before expanding it to businesses and ordinary consumers:
