Google releases Gemini 4 Argon, its new frontier model: Benchmarks, how to try it, and pricing

Google has officially rolled out Gemini 4 Argon, its latest frontier model, making it the top model in the company's Gemini 4 generation lineup. CEO Sundar Pichai announced the global availability on X, though the Tweet is currently inaccessible. Initially limited to trusted cyber defenders in Google's Fairwind Program, Argon is now accessible to Google AI Pro and Ultra subscribers through the Gemini app.

Argon is engineered for intricate, software engineering tasks, enterprise knowledge work such as legal and finance, and cybersecurity. In everyday conversations, it delivers concise responses with fewer hedging phrases, less repetition, and fewer clichés. When handling longer tasks, it provides frequent progress updates.

Google's internal evaluation places Argon ahead of its rivals on numerous benchmarks. On the Artificial Analysis Intelligence Index, Argon ties GPT-6 Astra at 53 with high reasoning enabled, one point ahead of GPT-6.1 Sol and the new Muse Wonder Max. However, it trails behind Claude Opus 5.5 and Claude Sonnet 5.5 by about five points. The index places Argon slightly behind Anthropic's Claude Opus 5.5 and Google's Claude Sonnet 5.5.

The Gemini app received several updates alongside Argon's release. Antigravity's agentic coding interface is now integrated within the Gemini app on macOS, enabling users to transition seamlessly between chatting and coding without switching applications. A new toggle in the main chat window allows users to switch from chat to task mode, enabling Gemini to work across Google apps and other services while the user is away from their computer. Pro and Ultra subscribers have access to these features.

Google also expanded Gemini app connections to over 40 apps, including Instacart, OpenTable, Adobe, and Shopify. Users can leverage these integrations for tasks like ordering groceries, booking restaurant tables, and designing or building an online business.

Benchmark performance showcases Argon's strengths and weaknesses. On the Vals Index, which accounts for finance, coding, legal, and tax work's share of U.S. GDP, Argon leads the pack. On AutomationBench, a test of end-to-end business tasks by Zapier, Argon scored 51.3%, nearly nine points ahead of the next model. In coding tests, Argon achieved a high of 77.9 percent on DeepSWE v1.1, Google's primary test of long software engineering tasks.

However, it trails GPT-6 Astra on FrontierSWE v2 and Terminal-Bench 4.0. Some figures for competitors seem inconsistent with Anthropic's reported results.