Gemini 4 Argon is Google's new top model for coding, finance and legal work. Google announced it on September 30 but gave it only to a small group of security teams, and said paid developers and subscribers to its top plan, Google AI Ultra, would follow. Argon has since appeared in Google's coding tool Antigravity, and Google's employees are testing a stronger version called Carbon, which one of them compared to Anthropic's Opus 5.5 for coding. Google has not given a date for the public release.
The launch price is the cheap half
At launch, Argon costs $2 per million input tokens and $10 per million output tokens. A token is a small piece of a word; input is what the model reads and output is what it writes. That rate is an introductory one, and afterwards the price becomes $4 and $20. Those are exactly the rates Anthropic charges for Opus 5.5, so the cheaper rate is a way to get people to try Argon, not a lasting advantage.
For example, a distributor has a tool that reads 500 supplier invoices a day. With instructions, each invoice is about 4,000 tokens going in, and the tool writes about 1,000 back. At the launch rate that costs $9 a day, and once the rate doubles it costs $18.
Now turn the effort up to high, so the model thinks for another 4,000 tokens on every invoice. On Google's Flash models, thinking is billed as output, and I have assumed Argon works the same way. At the later rate that adds $40 a day, and the total reaches $58. It is the same model reading the same invoices, and the daily cost is more than six times what the launch price suggested.
Two settings do most of the work
The first setting is how hard the model thinks. The Gemini app has replaced its single extended-thinking switch with low, medium and high, and free users, who can no longer pick a model at all, get the same three levels. The app now chooses the model for them, so effort is the one control that is left. A newer model can cost more at an unchanged price because it is built to think longer, and the way to bring the bill down is the same setting.
The second setting is how much the model reads at once. Antigravity's latest build lists three memory sizes for Argon: 256K tokens by default, 512K at about 1.3 times the quota per turn, and 900K at about 1.8 times. The largest option holds about 3.5 times as much for under twice the cost, so it is good value when one job needs a whole contract file or codebase in view, and a waste for a one-page email.
One more saving costs nothing. In the invoice example, about 3,000 of the 4,000 tokens are the same instructions every time, and Argon's cached input is priced 95% lower. Caching those instructions would take the input part of the later bill from $8 a day to about $2.30.
Build on the settings, not the model name
Nobody outside Google knows whether Carbon will arrive as an update to Argon or as a model of its own. A tool built around a model name has to be rebuilt when that happens, while a tool built around settings carries over.
Before the introductory rate ends, run ten of your real tasks at low, medium and high, and write down what each setting cost. Keep the cheapest one whose answers you would send out unchanged. When Carbon or any other new model arrives, repeat the test; it takes an afternoon.