Google launches Gemini 4 Argon, a top model for finance, law and cyber defense

What this is about

Google (Alphabet) has unveiled Gemini 4 Argon, its new flagship model. It replaces Gemini 3 as the most powerful model and follows a series of fast, cheap "Flash" models that appeared across the Google ecosystem over the course of the year. Google names programming, finance, law, cybersecurity and science as its focus areas. The sources give slightly different launch days (Tuesday and Wednesday); the reports come from Spanish-language outlets and rely on Google's statements.

Performance according to Google

According to Google's own evaluations, Argon is on a par with Claude Opus 5.5 (Anthropic) and GPT-6 Astra (OpenAI). Google stresses that its depth and accuracy on multi-step tasks let it beat other top models in several tests. It says the model leads the Vals Index, which measures performance in economically important areas such as finance, coding, law and tax. On complex financial research (Vals Finance Agent v2) and business process automation, it is said to be clearly ahead. It is also said to be strong at processing large amounts of information, searching extensive data, understanding long videos, reading complex charts and scientific research. One report also states that the model is level with GPT-6 Astra and Grok 4.7 in cybersecurity and beats newer Anthropic models on professional-performance indexes. That comes from a single source.

Weaknesses

Google concedes that Argon trails Claude Opus 5.5 and GPT-6 Astra at software development in the terminal and in very large repositories. On operating system control, GPT-6 Astra keeps its lead.

Conversational behavior

Google says Argon communicates clearly, keeps the user informed and follows style instructions better than any earlier Gemini model. It answers more directly, with less hedging, doesn't repeat itself unnecessarily and avoids clichés.

Internal use and rollout

According to one report, Google uses the system internally to optimize memory in its data centers, freeing up hundreds of terabytes of capacity without new hardware. Quantum researchers are said to have built it into their experiments. Product lead Tulsee Doshi told CNBC that the staged rollout builds more confidence and puts a model heavily trained in cyber defense into defenders' hands early:

This frontier-level model is of considerable significance for defenders.

(paraphrased translation). One report mentions that the rollout is happening in phases with cybersecurity partners and that Google is expanding safeguards in four areas. In addition, CEO Sundar Pichai signed a voluntary safety commitment with President Trump at the White House a day earlier.

Pricing and availability

You can use Argon in the Gemini app and in Google Search's AI Mode with Google AI Pro or Ultra. Developers will find it in the paid Gemini API through AI Studio and Android Studio, and businesses in Gemini for Enterprise, the API and Google Antigravity.

Safety

Google calls it its most comprehensive safety package to date. Capabilities and safeguards were reviewed with external partners, and testing covered cybersecurity, biological risks and autonomous behavior. One report points to OpenAI's problems, which led it to pause frontier training.

Our take

Argon makes Google a direct competitor at the top again, especially for enterprise tasks involving finance and law. All the comparison figures come from Google itself; independent tests, concrete benchmark numbers and details on the architecture are missing from the sources. The $2/$10 price is labeled a launch price, and the later price is unclear. For developers running terminal coding agents, Google's own statements point to drawbacks.