Gemini 4 Argon is late, but Google’s expertise may be an advantage

Google’s long-delayed flagship model, Gemini 4 Argon, shows that the cloud provider has been trailing competitors Anthropic and OpenAI in critical capabilities such as cybersecurity.

Introduced on Sept. 30, the foundation model can support long, multi-step tasks and perform at a high level in coding, reasoning, and in multiple modes, including images, text and video, as well as across diverse enterprise workflows. Google has secured Argon so that the model refuses risky requests, it said. It has also improved its internal activation mechanisms to detect misuse. The model is resistant to indirect prompt injections, in which a bad actor tries to hijack the model’s behavior, Google said

Argon is the tech giant’s first new top-of-the-line model since it released Gemini 3 Pro in November 2025. The new release indicates that Google is catching up with independent rivals Anthropic and OpenAI, both of which have released new flagship models in the past few months.

OpenAI released GPT-6 Astra in early September, while Anthropic released Claude Fable 5 and Claude Mythos earlier in the year. Like the cyber model Claude Mythos, which is accessible to trusted organizations under Project Glasswing, Google is first rolling out Argon to organizations in its cybersecurity program, Fairwind. OpenAI boasts a similar program, Daybreak, for GPT-5.6 Cyber.

Cybersecurity and Google tardiness

Google’s frontier model is the latest in a wave of AI models that specifically target security capabilities amid mounting concerns about AI safety. The model is capable of taking offensive cybersecurity measures.

“Cybersecurity has now become less of a unique offering; it has now become a staple for everyone,” said Lian Jye Su, an analyst at Omdia, a division of Informa TechTarget.

Google is trying not only to follow cybersecurity standards but also to “wrestle back what it has lost to OpenAI and Anthropic from an enterprise-grade agent or agentic AI standpoint,” he said.

With products such as Google Workspace, an integral part of many enterprise workflows, Google’s best option was to build a model with strong cyber capabilities, Su said.

“Whether we like it or not, we are going to be using a lot of AI, agentic AI workflow,” he said. “Having this sense of security being built into the model is quite important.”

Also, the tardy release might even help Google, said William McKeon-White, an analyst at Forrester Research.

He added that when Google recently admitted that Gemini had hacked three outside companies in May, it scored “a very odd win” because the vendor said Gemini was able to stop its own attacks.

Google revealed that during a cybersecurity evaluation by the lab Irregular, Gemini was asked to attack a fictional company that matched a real one. The AI model searched for the company on the internet and was able to guess the password for one company and a public code repository for two others. Gemini said it stopped the attack after logging into the networks and asserted no harm was caused to the companies.

While Gemini’s move is similar to Anthropic and OpenAI’s models escaping their testing environment and attacking an outside organization. Google showed that Gemini appeared better at recognizing its errors, McKeon-White said. Therefore, Google’s slower pace in releasing models suggests a need to be careful in releasing them to better manage their safety.

“Google was able to point to the models stopping the attack independently as proof of ‘this is why we have to do this carefully’ with the undercurrent of ‘and ours stopped,’” McKeon-White said.