Gemini 4 Argon is now available: who can use it, what it costs and what's new

When Google announced Gemini 4 Argon on September 30, the headline came with a catch: almost nobody could use it. The model went first to a set of trusted cyber defenders in Google's Fairwind Program, and businesses were told to wait. That wait is over. On October 13, Google made Argon available worldwide to developers and enterprises through its API, and to Google AI Pro and Ultra subscribers in the Gemini app and AI Mode in Search. ¹⁸

This guide covers what changed with the general release. If you want the full benchmark story, our first article on Argon walks through Google's 18-test table and the independent results from Artificial Analysis. Here the focus is practical: where you can use Argon today, what it will really cost once the introductory price ends, what Google's new robotics results mean, and a short trial plan for deciding whether it belongs in your work. ⁸⁹

What changed between September 30 and October 13

In September, Google described a staged rollout: trusted cyber defenders first, then paid API customers and Google AI Ultra subscribers, then everyone else, with no dates. The October release skipped most of that queue. Developers on the paid tier of the Gemini API can build with Argon now, Google Cloud customers can call it as a standard model, and both Pro and Ultra subscribers see it in the Gemini app. Google also put a date on the introductory price, which runs until November 30. ¹²⁴⁸

Where you can use Gemini 4 Argon

There are now five ways in, and which one fits depends on whether you write code. Businesses that want to build Argon into their own tools use the API, either through Google AI Studio or through Google Cloud, where the model ID is gemini-4-argon. Everyone else meets it as a choice in the Gemini app's model picker, or behind harder questions in AI Mode in Search, provided they pay for Google AI Pro or Ultra. ¹²⁴⁶

Two details matter for regulated businesses. On Google Cloud, Argon runs on the global endpoint and in the US and EU multi-regions, so a firm that needs client data processed in Europe has a supported option. And the free and lower Gemini app plans do not include Argon; a team that wants to try it without writing code needs Google AI Pro or Ultra seats. ⁵⁷

What it costs, and why November 30 matters

Google has kept the launch price of $2 per million input tokens and $10 per million output tokens, with cached input 95% cheaper, and has now said it lasts until November 30. After that, the standard price is $4 and $20, double the introductory rate. Any business case you build this month should use the standard price, because that is what you will pay from December. ¹⁴

Per-token prices also hide how many tokens Argon uses. In Artificial Analysis's testing of the preview, it averaged about 62,000 output tokens per index task, more than twice GPT-6 Astra's 27,000. That is why Argon's cost per task at the introductory price, $1.99, was only about 60% of Astra's $3.26 despite a far lower token price, and why simple doubling puts it at about $3.98 per task after November 30, slightly above Astra. GPT-6.1 Sol remains much cheaper at $0.72. ⁹

The benchmark picture has not changed

Google's release post points again to Argon ranking first on the Vals Index, which measures performance across finance, coding, legal and tax work, and its September table still shows the widest leads on legal, finance and business-workflow tests. Independently, Artificial Analysis put Argon level with GPT-6 Astra on its Intelligence Index at 53, with a hallucination rate of 15%, far below its rivals. Those independent results were measured on the preview; we will update this article if the released model scores differently. ¹⁸⁹

New: Gemini 4 Argon for robotics

The genuinely new material in this release is about robots. Google's developer team published results showing Argon reasoning about the physical world well enough to plan and check robot actions. On ERQA, a benchmark of embodied reasoning that asks questions about spatial relationships, actions and real-world scenes, Google reports that Argon scores 87% using reasoning alone, ahead of every other frontier model it tested. ³

Google's demonstration is deliberately ordinary. Argon directs a Boston Dynamics Spot robot to pick up a soda can, check its wrist camera, practice balancing on three legs in a physics simulator, step on the pedal of a trash bin, confirm its hand is centered over the open lid and drop the can in. Google presents this as zero-shot control, meaning no task-specific robot training: Argon planned the steps, tried them in simulation and checked each one with the robot's own camera. ³

The technique behind it is what Google calls Real2Sim. Argon builds a physically calibrated simulation twin of a scene from the robot's camera feeds, from video filmed from a person's point of view, or even from a sketch on paper. The robot then practices in the twin before acting for real. In one workflow Google describes, practicing in simulation raised the real robot's success rate from 56% to 75%. ³

For most businesses this is a signal, not a product. Nobody is going to buy Argon to run a warehouse this quarter. But the direction matters for manufacturing, logistics, inspection and facilities work: a general model that can turn a short video or a sketch into a practice environment lowers the cost of teaching a robot a new task, which has always been the expensive part. If your operations involve robots, this is worth a conversation with your integrator; if they do not, file it under what to watch in 2027.

Security: the guardrails story continues

Argon was trained to be highly capable at cybersecurity defense, and in September Google said that only trusted defenders and its own teams would get it without cyber guardrails, while it strengthened safeguards against misuse and prompt injection before a wider release. That wider release has now arrived. For businesses connecting Argon to email, documents or the web, Google's description of it as its most resilient model yet against prompt injection lowers the risk; it does not remove the need for a person to approve anything that sends, pays or deletes. ⁸

A two-week trial before November 30

The introductory price gives you about seven weeks of cheap testing. Two of them are enough to decide whether Argon earns a place in your work, provided you test on real tasks and judge the result on cost per finished task at the standard price. The steps below work whether you test through the API or by hand in the Gemini app.

What this means for your business

Two cautions carry over from the launch. First, budget on the standard price, because the introductory rate ends on November 30 and Argon uses a lot of output tokens per task. Second, the independent numbers so far come from the preview; the benchmark leads are real, but your own documents and workflows are the test that counts. ¹⁹