What happened
Google announced Gemini 4 Argon on September 30, but back then only a handful of cybersecurity teams could use it. Yesterday, October 13, it opened it to everyone who pays: the Gemini API, Google AI Studio, its Antigravity development environment, Gemini for business on web, desktop and mobile, and eligible Gemini Enterprise subscriptions. Google AI Pro and Ultra subscribers also get it in the Gemini app and in AI Mode in Search.
It is Google's big model, the one that competes with Claude Opus 5.5 and GPT-6 Astra. Google pitches it for complex, multi-step work: coding, finance, law and cybersecurity.
What interests us most is the price, because it has an expiry date.
The numbers worth knowing
Prices are per million tokens. A token is a piece of a word: a normal email is a few hundred, and a twenty-page contract can go over ten thousand.
- Until November 30, 2026: $2 input and $10 output. Cached input is 95% cheaper.
- From December 1: $4 input and $20 output. Exactly double.
- Claude Opus 5.5 costs $4 and $20. Claude Sonnet 5.5, $2 and $10. So Argon now costs what Sonnet does, and in December it will cost what Opus does.
- Claude Haiku 5.5, the small model we covered last week, costs $0.10 and $0.50. Twenty times less than discounted Argon.
- Google cites customers such as GitLab, Harvey and ServiceNow. ServiceNow says Argon improves its task success rate by around 20% over Gemini 3.8 Flash in its tests.
Why it matters
Five days ago we explained that automating simple tasks with AI now costs cents a month. Argon is the other end: an expensive model built for hard problems that is half price for six weeks.
There is a good reading and a risky one. The good one: if you had a project that needed a powerful model and dropped it because of cost, you can now test it cheaply. The risky one: if you build something on the launch price, in December the bill doubles without you changing anything.
One more warning. A model being new and expensive does not mean it is the best for your case. VentureBeat reports that Artificial Analysis ranks Argon's general intelligence below Claude Opus 5.5 and Sonnet 5.5, although it stands out in professional tasks. Rankings change every month. The only thing that counts is testing with your own data.
What it means for a small business
We did a quick calculation with the published prices. It is our own estimate and in dollars; what matters is the order of magnitude.
Imagine you want to review 200 contracts or tenders a month, with about 20,000 input tokens and 2,000 output tokens each. With Argon at launch price it would cost about $12 a month. From December, about $24. With Haiku 5.5, about $0.60.
Twelve or twenty-four dollars a month won't ruin anyone. The problem comes when someone uses the big model for everything. If you put Argon to sort thousands of emails or answer your website chat, you are paying for power you don't need.
Our rule, the same one we apply in the agents we build for clients:
- Big model for what is hard and infrequent: analysing a contract, reviewing an application's security, preparing a financial report from many sources.
- Small model for repetitive work: sorting, summarising, extracting data, answering frequent questions.
- Both together when needed: an agent that uses the small model day to day and only calls the big one when a case gets complicated.
What I would do this week
Six weeks is enough time to test calmly. This is what we would do:
- Pick one or two hard tasks that you do by hand today or that your current model gets wrong. Test them with Argon before November 30, while mistakes are cheap.
- Compare with what you already have. Same fifty real cases with Argon and with your current model. Look at quality, time and cost, not just the first nice result.
- Do the maths with the December price, not today's. If the project only pays off at $2 and $10, it doesn't pay off.
- Use caching. If you always send the same long instructions or the same base document, cached input costs 95% less.
- Build the project so you can switch models with one configuration line. We said it with Haiku and we say it again: competition between Google, Anthropic and OpenAI is your best insurance against price rises.
- If your company uses Gemini, also check what changes with Gems, which we covered in Google retires Gemini Gems.
What we think
We like Google competing on price at the high end, because it forces everyone else to move. What we don't like is the habit of launching with a temporary discount: it invites quick decisions and lets you discover the real price once you are already in.
For most small businesses we see, the bottleneck is not having the smartest model on the market. It is knowing which task to automate first and measuring whether it works. Argon is a good tool for a few specific tasks. There is no need to swap everything for it.
If you want us to look at it with you
We design and maintain AI agents and automations for small businesses, and in each case we choose which model fits and how much it will cost. If you would rather have someone make these technical decisions on an ongoing basis, we do it as your fractional CTO.
Book a thirty-minute call. We look at the tasks you have in mind, whether any of them justifies a model like Argon and how much it would cost you at the December price. No commitment.
Sources
This article is based on the following publications, consulted on October 14, 2026:
Frequently asked questions
It is Google's new flagship model, announced on September 30, 2026 and opened to developers and businesses on October 13. It is designed for complex multi-step tasks in coding, finance, law and cybersecurity.
Until November 30, 2026 it costs $2 per million input tokens and $10 per million output tokens, with a 95% discount on cached input. From December 1 it rises to $4 and $20.
In the Gemini API, Google AI Studio, Google Antigravity, Gemini for business and eligible Gemini Enterprise subscriptions. Google AI Pro and Ultra subscribers also get it in the Gemini app and in AI Mode in Search.
At launch price it costs the same as Claude Sonnet 5.5. From December it will cost the same as Claude Opus 5.5. Compared with small models like Claude Haiku 5.5, at $0.10 and $0.50, it is about twenty times more expensive even with the discount.
For difficult, low-volume tasks such as reviewing contracts or auditing code, it can. For repetitive tasks like sorting emails or answering frequent questions, a small model does the job for a fraction of the cost.
You can test with it, but do the maths with the December price. If the project is only profitable at $2 and $10, it will stop being profitable in December because the cost doubles.