Google Launches Gemini 4 Argon Worldwide as Its New Flagship Frontier AI Model

Google has officially launched Gemini 4 Argon worldwide, designating it as the company's most capable flagship AI model to date. Replacing the previous Gemini Pro tier as Google's frontier series, Argon is now available across enterprise channels, consumer apps, and developer platforms.

The model is accessible via API in Google AI Studio and the Gemini platform. In terms of user tiers, Google AI Ultra subscribers can access Argon inside the Google Antigravity workspace, while both AI Ultra and AI Pro users can utilize the model through the Gemini application and Google Search’s AI mode. However, free users and paying Plus subscribers will not have access to Argon immediately; under the new tier system, they remain on weaker Flash models until updated Gemini 4 Flash variants arrive.

Internally, Google is already testing a follow-up model codenamed "Carbon," which reportedly targets even higher software engineering and coding capabilities.

Aggressive Pricing and Benchmark Landscape

Google has established competitive launch pricing for Gemini 4 Argon. The promotional API rate sits at $2 per million input tokens and $10 per million output tokens, alongside a 95 percent discount on cached inputs ($0.10 per million tokens). Standard regular pricing will later rise to $4 for inputs, $20 for outputs, and $0.20 for cache reads.

By comparison, Anthropic’s Claude Opus 5.5 charges $4 per million input tokens and $20 per million output tokens, while GPT-6 Astra and Claude Fable 5.1 are priced at $10 for inputs and $50 for outputs.

In terms of performance rankings, Argon captured the top position on the Vals Index. However, it trails Anthropic’s Claude Opus 5.5 in independent aggregate evaluations, and sits behind Meta’s newly launched Muse Wonder Full.

Internal Engineering and Coding Gains

Google reports substantial productivity dividends from deploying Argon internally. Software engineers using the model saw agent-driven code submissions rise by over 35 percent, while rollback rates decreased by nearly one-third.

A prominent real-world deployment took place within the Chromium project. Argon assisted the Chrome engineering team in executing an architectural refactor across hundreds of files within its 40-million-line codebase—compressing months of scheduled refactoring work into a few weeks. The structural update allows web pages to transition across tabs, floating windows, and sidebars without triggering reloads.

In automated testing with Google's open-source Agent Development Kit, Argon operated autonomously overnight to find system optimizations, reducing session preparation time by 40 percent and cutting cold import delays by 21 percent. Additionally, upstream pull requests show that open Gemma model inference speed on Nvidia A100 GPUs via vLLM increased by a factor of 2.85, rising from roughly 37,900 to more than 108,000 tokens per second.

Beyond standard code, Argon demonstrates direct browser-based graphics and 3D capabilities. Google showcased a complete raytracing engine written in pure Python without external dependencies, as well as an autonomous workflow in Antigravity where Argon used the Model Context Protocol (MCP) to drive Blender, constructing a gothic architectural scene with audio-reactive lighting starting from a single prompt image.

Alignment, Safety Benchmarks, and Research Access

Alongside the launch, Google announced research and security grant initiatives:

• Academic Access: Up to 10,000 researchers can apply for six months of free access to the highest tier of Google Antigravity on a rolling basis.
• Fairwind Program: Specialized access granting defensive cybersecurity researchers access to Argon without typical cyber safeguards, which Google plans to expand to hundreds of partners.

On alignment and safety, Google claims Argon has its strongest safety architecture to date. External red teaming in biological domains bypassed less than 1 percent of safeguards, while general adversarial violations were cut roughly in half compared to prior Gemini generations. Unauthorized agent shortcuts dropped to 2 percent, monitored in real time by an automated misalignment tracking system.

On the Gray Swan Indirect Prompt Injection benchmark, Argon achieved an attack success rate of 0.7 percent across 15 adaptive attempts, comparing favorably to prior recorded marks of 4.8 percent for Claude Opus 5 and 8.5 percent for GPT-6 Astra.