Executing complex, multi-step tasks... "Gemini 4 Argon" revealed

Google has unveiled its advanced artificial intelligence model, Gemini 4 Argon, and made it available to developers and companies across the globe, in a move aimed at reclaiming its competitive standing against OpenAI and Anthropic. This launch follows months of delays and significant leadership restructuring within DeepMind.

Google, a subsidiary of Alphabet, stated that the new model excels in what it terms “deep work,” executing complex, multi-step tasks in sectors such as software development, finance, law, cybersecurity, and science.

At launch, the company set the API pricing at $2 per million input tokens and $10 per million output tokens, with prices set to double after November 30.

The release comes after a challenging period for Google in the AI field. In September, the company announced it would no longer proceed with the launch of Gemini 3.5 Pro, a version that CEO Sundar Pichai had previously promised to release in June.

Additionally, the company undertook a comprehensive reorganization at DeepMind following the resignation of co-founder and CEO Demis Hassabis, as well as the departure of several leaders from the Gemini team.

Argon is available to subscription customers through Google’s tools designed for building AI applications. Gemini AI Pro and Ultra subscribers can access it via the Gemini app and AI Mode in Search, while the coding tool Antigravity is exclusively available to Ultra subscribers.

Google noted that teams using Argon within Antigravity recorded an increase of over 35% in AI agent code submissions, accompanied by a reduction in code reverts of approximately one-third.

Early testers of the model include Deloitte, LexisNexis, Ambiance Health Care, and Isomorphic Labs.

In a parallel development, Google announced it would open its “Friend” program to hundreds of new partners. This program allows cybersecurity specialists, after passing an audit, to obtain an unrestricted version of Argon, free from built-in safety constraints.

In September, the company had previously indicated that it was making the model available to selected partners in the cybersecurity sector.