Google Launches Gemini 4 Argon, Its Most Advanced AI Model Yet

Google has officially launched Gemini 4 Argon worldwide, positioning it as the company's most advanced AI model available for developers, businesses, and consumers.

Gemini 4 Argon can be accessed through the API within AI Studio and the Gemini platform. Additionally, subscribers of Google AI Ultra can utilize Argon in the Antigravity tool. However, accessibility varies: free users and paying Plus subscribers are currently restricted to the less capable Flash models under a new tier system. Argon is effectively replacing the former Gemini Pro class as Google's premier series of AI models, indicating a shift in their strategic focus toward higher-performing variants with more features. There are reports of a potential upcoming successor, referred to as "Carbon," which is expected to excel particularly in coding tasks.

The pricing strategy is crucial for understanding Gemini 4 Argon’s place in the AI landscape. Initial rates are set at $2 per million input tokens and $10 per million output tokens, with a notable 95 percent discount on cached inputs. Notably, Argon currently ranks highest on the Vals Index but still lags behind competitors, such as Anthropic's Claude Opus 5.5 and Meta's Muse Wonder Full, with the latter having launched just a day before Argon. The table below summarizes the pricing for different models:

*Pricing for cache reads is estimated as Google notes a 95 percent reduction compared to standard input token costs.

Performance Enhancements with Argon

Google has provided some internal metrics that indicate significant improvements resulting from Argon's launch, though these figures are self-reported. According to the company, engineers utilizing Argon experienced an increase of over 35 percent in agent-driven code submissions, while code rollbacks dropped by almost a third.

In specific applications like the Chromium project, Argon facilitated the Chrome team’s architectural refactor across a massive 40-million-line codebase, completing tasks that would typically require months in just a few weeks. This rapid progress enhances user experiences, allowing for web pages to transition seamlessly among tabs without unnecessary reloads.

Utilizing the open-source Agent Development Kit, Argon has also autonomously optimized several parameters, reducing session preparation time by 40 percent and improving cold import speeds by 21 percent. Notably, inference performance surged from approximately 37,900 to an impressive 108,000 tokens per second on A100 systems using vLLM. The results of these performance increases are documented in upstream pull requests, evidencing real improvements in diverse applications.

In another fascinating application, Argon can autonomously generate intricate 3D models, games, and animations purely via coding, a capability evidenced by a demo showcasing a raytracing engine crafted entirely in Python and devoid of external libraries. The complete source code for this demo is publicly available for further examination.

Expanded Access and Safety Measures

As part of its commitment to research, Google is offering free six-month access to the highest tier of Google Antigravity for 10,000 researchers. This initiative will be merit-based, assessed according to eligibility and availability, with applications currently being accepted. Additionally, cybersecurity defenders have a separate opportunity to apply for the Fairwind program, which aims to provide unencumbered access to Argon.

On the safety front, Google claims that Argon's architecture represents its best safety features to date. Independent assessments for external threats in sensitive domains demonstrated a bypass rate of under 1 percent on tested safety checks. Additionally, incidents of misalignment with the agent's objectives fell dramatically to just 2 percent. An automated monitor evaluates reasoning and actions in real time, helping maintain a secure operating environment. Moreover, Argon achieved a mere 0.7 percent success rate for the Gray Swan Indirect Prompt Injection test, outperforming other models like GPT-6 Astra and Claude Opus 5, which scored significantly higher on comparable benchmarks.

As developments unfold, Google appears poised to maintain a competitive edge in AI while continuing to enhance both performance and safety in future models, such as the anticipated Carbon.