Google is bringing its flagship AI model, Gemini 4 Argon, to users around the world. After nearly three weeks of limited access for cybersecurity defense partners, the model was officially broadened on Tuesday to serve general users, developers, and enterprises. Google aims to use its strengths in coding, complex reasoning, and cybersecurity to further solidify its position in the race at the AI frontier.
On Tuesday, October 13, U.S. Eastern Time, Google announced that Gemini 4 Argon has launched globally. Subscribers to Google AI Pro and Ultra can experience the model in the Gemini app and in AI Mode within Google Search; developers and enterprise users can call it through the paid Gemini API. Ultra subscribers can also use the model in Google AI Studio and Antigravity, while enterprise customers can access it through Gemini enterprise services and Antigravity.
Google CEO Sundar Pichai announced the official launch of the new model on social media. Google DeepMind said Argon is designed to solve complex problems from start to finish, delivering frontier capabilities in coding, enterprise workflows, and cybersecurity defense. Logan Kilpatrick, product lead for Google AI Studio, said the model had been tested by more than 100,000 people inside Google over the past month, and the R&D team invested about six months in this release.
The global rollout marks a new phase for Google's previously cautious release strategy. As the potential cybersecurity risks of increasingly capable models sparked widespread discussion across the tech world, Google chose to let professional defenders use Argon first and then gradually expand access after relevant organizations had a chance to patch vulnerabilities and fix system flaws uncovered by the model.
From Targeted Cybersecurity Access to a Global Launch, Google Expands Model Availability
When Google first released Gemini 4 Argon on September 30, it did not immediately open it to the public, instead providing access to vetted cybersecurity defenders through the Fairwind program.
Google said at the time that, given the growing cybersecurity capabilities of frontier models, a phased release strategy was needed so defenders would have time to identify and patch vulnerabilities while security safeguards continued to improve.
This expansion marks the beginning of Google pushing the capabilities of its flagship model into broader consumer and commercial scenarios.
Still, access is not without barriers: ordinary users need a Google AI Pro or Ultra subscription; developers connect through the paid Gemini API; and Google AI Studio and Antigravity are open to Ultra subscribers. Google also allows eligible scientists and researchers to apply for free use of Antigravity's highest usage tier, further promoting the model's application in scientific research.
Output Limit Raised to 1 Million Tokens, Targeting Complex Coding and Long-Horizon Tasks
Argon's core positioning is not simply answering questions, but handling complex tasks that require long reasoning and multi-step execution.
Google said the new model is optimized for large-scale software engineering, enterprise knowledge work, and cybersecurity defense, and can be used for large codebase migrations, complex engineering problems, legal and financial research, and other tasks. Compared with traditional one-question, one-answer interactions, such work often requires the model to maintain contextual coherence over a longer period and repeatedly plan, execute, and check results.
On technical specifications, one major breakthrough for Argon is raising the maximum output token limit from the previous 64,000 to 1 million. This means that, under applicable background execution modes, the model can generate longer continuous output, providing more room for large code modifications, complex research reports, and multi-step agent workflows. It is worth noting that 1 million tokens is the maximum output capability and does not mean every call will generate content of that length.
Kilpatrick, product lead for Google AI Studio, said Argon has made progress in coding ability, knowledge reserves, and sustained task handling. Google had also previously used the model for internal engineering work, and the latest disclosed cases include automated refactoring and optimization of large codebases, reflecting its positioning for real development processes.
Beyond software development, Google also released an introduction on Tuesday to Argon's applications in robotics.
The model can combine multi-view video, three-dimensional geometric information, physical simulation, and the robot's own state data to reason about the physical relationships between objects and connect visual understanding, code generation, and robot control. Google hopes to use this to advance the model's application in real physical environments.
As shown below, Argon controls a Boston Dynamics Spot quadruped robot as it learns a new skill: stepping on a pedal-operated trash can to discard a soda can. Before actually pressing the pedal and throwing the soda can into the open bin, Argon first creates a MuJoCo articulated digital twin of the pedal-operated trash can to practice the relevant motions.
Leading 13 of 19 Benchmarks Individually, Squaring Off Against OpenAI and Anthropic
At its initial release on September 30, Google published Gemini 4 Argon's performance across 19 benchmarks. According to the results Google disclosed at the time, the version with the highest thinking tier achieved the top individual score in 13 of those tests, with another tied for first.
Those results show that Argon particularly emphasizes knowledge-intensive tasks, long-context reasoning, and complex coding capabilities. Its main competitors include frontier models such as OpenAI's GPT-6 Astra and Anthropic's Claude Fable 5.1. The results provide a basis for Google's return to competition among the world's top models, but leading in benchmarks does not mean the model outperforms competitors in all tasks and real-world applications.
This global launch also moves these capabilities from a restricted security testing environment into broader user scenarios. For Google, the key going forward is not only benchmark performance, but also developer adoption, the actual quality of completing complex tasks, and whether the model can continue to create value in enterprise workflows.
Promotional API Pricing Below Major Rivals, Prices Double After End of November
Pricing is another important lever for Argon in competing for developers and enterprise customers.
Google has set limited-time promotional pricing for the publicly available Argon, lasting through November 30, 2026, at $2 per million input tokens and $10 per million output tokens. After the promotional period ends, prices will rise to $4 and $20 respectively, doubling for both input and output.
The discount for cached input tokens reaches 95%. Based on promotional-period input pricing, the cost of cached input tokens is about $0.10 per million; after the price adjustment, it is about $0.20 per million. For applications that need to repeatedly call the same background materials, process long contexts, or execute multi-turn tasks, caching is expected to reduce some usage costs.
According to previously disclosed standard API prices for competitors, GPT-6 Astra and Claude Fable 5.1 charge $10 per million input tokens and $50 per million output tokens, both five times Argon's promotional pricing. Claude Opus 5.5's corresponding prices are $4 and $20, comparable to Argon's pricing after the promotional period ends. Billing rules, caching policies, and task performance differ across models, and actual costs still depend on the specific workload.
Overall, Google's launch sends two signals: on one hand, Gemini 4 Argon is moving from limited testing in cybersecurity into broader commercial applications; on the other hand, Google is trying to use long context, complex task execution capabilities, and competitive early pricing to attract developers to incorporate it into actual workflows. Whether the model can further expand its market share still remains to be tested by future user adoption and real business performance.
Comments