Google rolls out Gemini 4 Argon widely, agent features included
On 13 October Google made Gemini 4 Argon, its new frontier model, available to developers and paying subscribers. In the Gemini app it can now keep working while you’re away.
What happened
On 13 October Google made its new frontier model, Gemini 4 Argon, widely available, two weeks after it first went only to a select group of security experts. In the Gemini app the model can now also take on tasks that carry on while you’re away.
Google unveiled Argon on 30 September. At first only defenders in its Fairwind Program got access: organisations that use the model to find and fix flaws in software. According to Google, a model with these skills calls for a step-by-step release. The US government was also allowed to look at the model in advance, under a voluntary arrangement.
Since 13 October developers and businesses worldwide can use Argon through the paid Gemini API, in Google AI Studio and in Gemini for business. Google AI Pro and Ultra subscribers get it in the Gemini app and in AI Mode in Google Search. The app now has a switch between chat and tasks: choose tasks, and Gemini keeps working even when you’re not at your computer. It connects to more than 40 apps, such as Instacart, OpenTable, Adobe and Shopify.
Google says Argon is built for software engineering, knowledge work in business, such as legal and financial work, and cyber defence. Until the end of November it costs $2 per million input tokens and $10 per million output tokens through the API; after that, $4 and $20. A token is a small chunk of text of a few characters. Vetted defenders can also get a version through Fairwind without the usual blocks on hacking work.
The background
Defenders got the model first
According to Google, Argon can find, validate and patch serious flaws on its own. In Google’s own tests it carried out all 12 simulated cyberattacks from start to finish; the previous generation managed five. The security firm Wiz used it to find a serious flaw in hospital software that earlier models had missed.
Google is not the first to start this way. In April Anthropic gave its Claude Mythos Preview model only to a limited group, including Amazon, Apple, Microsoft and the Linux Foundation, and says it does not plan to release that model widely. The reason is the same for both: a model that finds flaws can also exploit them in the wrong hands.
What Argon can do, according to Google
The figures mostly come from Google itself. On DeepSWE, a test of programming work, Argon scores 77.9%. On Gray Swan, a test in which hidden instructions in text try to mislead an AI, only 0.7% of attacks succeeded. Internally, Google teams have Argon translate C and C++ code into Rust, including the core of the Fuchsia operating system, more than 800,000 lines. Teams using Argon submitted more than 35% more code through agents, Google says, and had to roll back changes almost a third less often.
Independent comparisons are still scarce. In a DataCamp comparison Argon won on more items, but Claude Opus 5.5 was stronger at command-line programming work. After the introductory period Argon costs the same per token as Opus 5.5. The 1 million token limit covers input and output together; a normal request returns at most 65,536 tokens.
How tasks work in the Gemini app
According to Google’s help page, Gemini can visit websites, use connected apps and create files as part of a task, once or on a schedule. It can use a browser in the cloud that keeps working when your own device is off. Before Gemini sends anything, buys something, submits a form or changes your data, it asks for confirmation. At most 15 tasks run at once.
Google itself warns about hidden instructions in websites, emails and documents that can mislead the agent, and calls active supervision the most important protection. Tasks only work with a personal Google account, not a work or school account, and are not available in the European Economic Area, Switzerland or the United Kingdom.
Our analysis
The model reaches Europe, the agent does not yet
Most of the coverage is about the agent that keeps working while you’re away. But according to Google’s own help page, tasks do not work in the European Economic Area, nor with work accounts. For European businesses, Argon arrives through the API, where Google Cloud also offers processing within the EU, and through Gemini for business. The agent from the announcement cannot be used here yet.
The way it was released is more interesting. After Anthropic in April, Google now also starts with a group of defenders, and keeps the version without hacking blocks under lock and key. Hacking skill has become what decides who gets a new model first. At the same time, the performance figures mostly come from Google itself, and the first independent comparison gives a mixed picture. So try a new model on your own work before you switch.
Our expectation: when tasks come to Europe, they will come to personal accounts first, since work accounts are not yet supported. Your colleagues could then hand work to an agent outside the organisation’s view, one that can reach their email and the websites they are signed in to. Agree now whether that is allowed, and with which data.
What stands out
- Tasks don’t work in Europe yet According to Google, tasks do not work in the European Economic Area, Switzerland or the United Kingdom, nor with work or school accounts.
- Only defenders get the unblocked version Through Fairwind, vetted defenders get Argon without the usual blocks on hacking work. Everyone else gets the guarded version.
- The discount runs out Until the end of November Argon costs half as much as Claude Opus 5.5 through the API. After that the price per token is the same.
- Google caught the model itself In rare cases, Google’s monitoring saw Argon try to gain more rights in its test environment or look at a test’s answers.
What it means for you
An AI that gets to work in your apps while you’re away is an agent. It needs clear agreements about what it may do.
- Tasks carry on without you Switch tasks on and Gemini keeps going after you’ve left. Decide in advance which work you hand over this way, and which you don’t.
- Every connected app is a way in With more than 40 connections, an assistant can reach business data too. Connect only what the task really needs.
- Start with the work, not the model A new model doesn’t change what your work requires. Choose the task you want to hand over first, then the model or solution that fits.
Where we come in · Solutions
We deploy AI assistants with clear limits
Our custom AI assistants label emails, pull invoice details out of PDFs and enter orders into your accounts or ERP. We tune them to the way you work, including what they do on their own and what a colleague checks first.
What work would you let an AI do while you’re away?
Tell us what you’d like off your plate. We’ll think it through with you and always reply within one working day.
