Google unveils new AI model

On 30 September, Google introduced Gemini 4 Argon, its new high-end AI model designed for complex tasks across multiple fields.

Deployment timeline and pricing

The model began global deployment on 13 October 2026 for developers, businesses, and paying subscribers of Google AI. Individual users can access Argon through the Gemini application and the AI Mode of the Google search engine, but only if they are Google AI Pro or Ultra subscribers.

Access restrictions and new partnerships

The free version of Gemini and the AI Plus offer, priced at 4.99 euros per month, does not have access to Argon. In France, Google charges 21.99 euros per month for AI Pro and AI Ultra starting from 99.99 euros. Previously, access was limited to highly selected cybersecurity teams in the Fairwind Program, but the company has now expanded the program to hundreds of new partners.

Scientific and research access

Additionally, 10,000 scientists will be able to use the highest level of Antigravity for six months for free, upon application. Antigravity is the development platform where AI agents write and test code.

Capabilities and performance

Argon is designed for long and complex tasks in code, finance, law, cybersecurity, or scientific research. It can generate up to 1 million tokens in a single response via the API, allowing it to work on large projects without interruption.

User interface and features

In the Gemini application, users can expect more direct responses with fewer repetitions. A button also allows switching from conversation to tasks, which Gemini can perform even when the user is not present.

Usage in development and pricing

Google Antigravity offers Argon only to Ultra subscribers. Android developers can use the model in Android Studio via Antigravity in the test version Canary or via the paid Gemini API. Developers using the API pay based on their consumption, measured in tokens.

Launch pricing and future changes

Until 30 November 2026, Google offers a launch price: 2 dollars per million input tokens and 10 dollars per million output tokens. After that, the price doubles to 4 and 20 dollars. Tokens already sent and stored in cache cost 95% less.

Performance improvements

Google claims that Argon's code submissions generated by AI agents have increased by more than 35% in recent weeks within its own teams, while the number of rollbacks has dropped by nearly a third. The Chrome team used Argon to reorganize a folder of about 6 million lines of code, out of the 40 million that Chromium contains. The work took a few weeks instead of several months.

Research benchmarks and comparisons

In terms of research, Google highlights that Argon achieves 87% on ERQA, a test measuring a model's ability to solve physical problems in a 3D environment. During an internal test, Argon watched a video of a person playing cornhole and then wrote the simulation code that allowed Apollo, the humanoid robot from Apptronik, to successfully make seven consecutive throws.

Benchmark rankings and independent evaluations

On benchmarks, Google highlights that Argon ranks first on the Vals Index, achieving 68.9%, compared to 67% for Claude Opus 5.5. However, the independent firm Artificial Analysis ranked it eighth overall at the beginning of October, five points behind Claude Opus 5.5.

Partnerships and use cases

Google has tested Argon with early access partners, including Deloitte, LexisNexis, Ambience Healthcare, and Isomorphic Labs. Their use cases include document synthesis, reasoning on large codebases, or analysis of longitudinal data. In cybersecurity, the company Wiz integrated the model into its initiative "Scan for Good," which protected hundreds of thousands of confidential medical files.

Cost comparisons and future pricing

Google states that Argon costs five times less than GPT-6 Astra (OpenAI) and Claude Fable 5.1 (Anthropic), and twice as little as Claude Opus 5.5. However, starting December 1, 2026, the prices will increase to $4 and $20, reaching the same level as Claude Opus 5.5. The question remains whether this model marks the end of Google's "Pro" line.