Google finally has an answer to rival the likes of OpenAI’s GPT-6 Astra and Anthropic’s Claude Mythos at the peak of AI intelligence. Earlier today, the company announced Gemini 4 Argon. Just like rival AI models from its competitors, Google says that it’s currently working with the US government to offer them access, while also working with a small batch of testers before Gemini 4 Argon is widely released.
Gemini 4 Argon posts big wins across key benchmarks
Based on benchmark figures released by Google, Gemini 4 Argon overwhelmingly beats GPT-6 Astra and Claude Fable 5.1 on knowledge work. It also races ahead of them on two out of four coding benchmarks, and puts out a similar win tally on science and maths-based tests. The former category is a huge win because, in the past few months, OpenAI has raced ahead with ChatGPT Codex, while Claude remains a preferred choice in the developer community for coding-related workflows.
Gemini 4 Argon also beats its competitors at the frontier of AI intelligence when it comes to producing long-context work, cybersecurity intelligence, and tasks that require multimodal input and output. Google claims that Argon can also handle cybersecurity-related chores with high efficiency, adding that it can “autonomously find, validate, and patch critical software vulnerabilities.”
Google is keeping Argon on a short leash for now
But it’s a frontier AI model at the end of the day, and as such, Google is playing it pretty safe. Over the past few months, ultra-powerful AI agents developed by OpenAI and Anthropic have engaged in concerning behavior, utilizing unexpectedly dangerous tactics to break into the systems of other companies. Such behavior is commonly referred to as misalignment in the AI industry, and it remains one of the biggest challenges of the field.
To address that and prevent Gemini 4 Argon from stepping outside of its test boundaries, Google says it is deploying measures to contain misalignment incidents while monitoring its chain of thought and actions, so that it can be stopped when needed. Google won’t say when exactly it will release Gemini 4 Argon, but it will only happen once it has finished testing with US government agencies and a small pool of trusted partners.
Once the vetting is over, rollout of Gemini for Argon will begin with paid API customers and folks who pay for the $200 per month Google AI Ultra subscription. Once that happens, the model will be released to developers, enterprising customers, and general paid consumers, likely in a diluted form.