Gemini 4 Argon arrives on the Google app and Google Search; learn more

Google began making Gemini 4 Argon available to more users this Tuesday (13). The artificial intelligence (AI) model, announced at the end of September, is available to subscribers of the Google AI Pro and AI Ultra plans in the Gemini app and in Google Search’s AI Mode. The feature is also coming to the Antigravity development platform, but in that case, access requires an Ultra subscription.

According to Google, Argon was developed to handle complex reasoning tasks and improve performance across different types of work. The company also says the model responds more directly, reduces repetitions and clichés, and follows style instructions more consistently. There is still no confirmation regarding availability for free users or subscribers of other plans.

Gemini gains new features with Argon

In the Gemini app, the model is now included among the options available to Pro and Ultra plan subscribers. Previously, initial access was restricted to the Ultra plan.The list of models available to these users includes 3.5 Flash-Lite, 3.8 Flash, 3.1 Pro, and Argon 4.

Google is also modifying the app’s interface to make it easier to switch between conversations and tasks. The tasks option, formerly called Spark, will have its own button in the main chat window.

Another new feature involves the macOS version of Gemini. The app will provide direct access to the Antigravity agent programming interface, allowing users to switch between conversations and coding activities to develop apps, web experiences, and other projects.

According to the company, the model can also share progress updates on longer-running tasks more frequently. The goal is to keep users informed while the system works on requests that take more time.

Google highlights results from internal projects

Google says it is already using Argon on internal projects. One example involves the Chrome browser: the model helped separate core web page processing from a legacy folder containing about 6 million lines of code, within a Chromium codebase totaling 40 million lines.

According to the company, the work involved changes to hundreds of files and made it possible to complete an architectural refactoring in weeks that would have taken months. The goal is to enable moving open pages between tabs, side panels, and pop-out windows without reloading, while also eliminating a category of browser failures.

Another cited case involves Gemma 4 E2B.With the use of vLLM on an A100 GPU, Google reports achieving a 2.85x speedup in end-to-end inference, increasing throughput from approximately 37,900 to over 108,000 tokens per second. The company says the result was achieved in a matter of days, without any loss of quality and with 100% token equivalence.

On YouTube, Argon helped halve the CPU usage of the comments interface. The analysis found that redundant emoji rendering accounted for nearly 45% of the server overhead. After 12 optimizations, server-side comment-loading latency fell by 30 milliseconds, while client-side end-to-end latency decreased by 20 milliseconds.

Gemini 4 Argon pricing and availability

In addition to the Gemini app and AI Mode in Google Search, the model is available through the Gemini API, with a promotional price of US$2 per million input tokens and US$10 per million output tokens through November 30, 2026.

After the promotional period ends, the prices will increase to US$4 per million input tokens and US$20 per million output tokens. Access to the model in Antigravity for desktop, the Antigravity command-line interface (CLI), AI Studio, and Android Studio Canary requires a Google AI Ultra plan.

Google has also opened registration for a program aimed at the scientific community.The initiative plans to offer the highest usage tier of Antigravity free of charge for six months to 10,000 scientists.

The company also states that Argon delivers competitive results in areas such as programming, cybersecurity, science, finance, and law. In benchmark evaluations, Google says the model outperforms competitors such as GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5 in different categories.

(empty)