← All articles

Mikhail Polyakov

Google releases Gemini 4 Argon: what is claimed, what it costs and how to test the model on your own tasks

Level: BeginnerReading time: 6 min

Topics: AI, Small business

A two-pan balance with a stack of smooth coins on one pan and a glowing blue faceted crystal on the other

Short answer: on 13 October 2026 Google began giving subscribers the Gemini 4 Argon model and said it is on a par with the best models in the industry. The claims of beating competitors in tests are Google's own words, and the article contains no independent results. The price stands out most: for the introductory period it is 2 dollars per million input tokens and 10 dollars per million output tokens, and at the end of November it will double. Below is what exactly was announced, what it means in plain words, and how to test the model on your own tasks before you change provider.

What was announced

There is one source: Robert Hart's article in The Verge dated 13 October 2026. This is a retelling; I did not read Google's own page.

  • What the model is. According to Google, Gemini 4 Argon delivers "frontier reasoning and improved performance across complex workflows", and also fewer hedges, repetitions and cliches, which chatbots are known for.
  • Who it is compared with. Google claims that Argon outperforms GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 in a number of tests: in programming, cybersecurity, science, finance and law. This is the company's claim; the article has no independent results.
  • Where it is already used inside Google. According to the company, the model finds and patches vulnerabilities, halved the computing power needed for the YouTube comments section, and helped debug the drone physics simulator for the delivery service Wing, which is part of Alphabet.
  • Who can use it. Since 13 October the model is being rolled out to subscribers of the more expensive Pro and Ultra tiers in the Gemini app and in AI mode in Search, and to Ultra subscribers on the development platform Antigravity. Whether and when it will reach Plus subscribers and free users is not stated in the article.
  • How it was released. Before the public launch, Argon was given to a limited group of "trusted cyber defenders" to test for risks. The author writes that other recent frontier models are released the same way.
  • Context. This is Google's first major frontier model since Gemini 3 in late 2025. Over that time, the author writes, models from Anthropic and OpenAI have moved well ahead. Argon came out the day after Meta announced its own model, Muse Wonder Full.

What the price means

Models are sold by "tokens": small pieces of text, roughly a word or part of a word. Input tokens are what you sent to the model, output tokens are what it returned.

  • Launch price. 2 dollars per million input tokens and 10 dollars per million output tokens. According to the author, this is about a fifth of the price of OpenAI's GPT-6 Astra and Anthropic's Fable 5 models.
  • What comes next. The introductory period ends at the end of November, and prices will double. The author notes that even then Argon will remain noticeably cheaper than its rivals per token.
  • What the price applies to. The article does not say which access method this price applies to or how it relates to the Pro and Ultra subscriptions. A price per token is not the cost of a subscription.
  • The author's caveat. A low price per token does not necessarily mean a low cost of work, because different models spend a different number of tokens on the same task.

Three checks for a company

These are my suggestions, not data from the article.

  1. Test on your own tasks, not on tables. A company's claims of winning in tests speak about those tests, not about your letters, contracts and spreadsheets.
    • Who does it: the person responsible for AI work in the company, together with an employee who knows these tasks.
    • How to check: take 20 real, typical requests that you have already handled. Give them to your current model and to Argon. An employee, without knowing which answer belongs to which model, marks each one "suitable" or "not suitable". Remove personal and commercial data from the requests until you have found out how the provider handles it.
  2. Count the cost of a task, not of a token. The price per token is half of the calculation.
    • Who does it: the same employee together with an accountant or a manager.
    • How to check: if you are offered payment per token, record on the same 20 requests how many tokens each model used and multiply by the price. For Argon calculate twice: at the introductory price and at the doubled price. The article does not say which access method this price applies to, so ask the provider which tariff you would actually pay.
  3. Do not tie yourself to one model. The balance of power may change, as the author of the article himself reminds us.
    • Who does it: the person who configures the system.
    • How to check: on a test copy that is not reachable from the internet, replace the model with another one and make sure the work continues. If the system broke, you have found a hard dependency that is worth cutting before you choose a provider.

What this costs

The calculation uses conditional numbers; substitute your own. Suppose the check takes 3 hours of an employee at 1500 roubles and one hour of a manager at 3000 roubles. Suppose also that you are offered to pay at the prices from the article, that you send the model 5 million input tokens and 1 million output tokens a month, and that it spends as many tokens as Argon. At the introductory price that is 5 × 2 + 1 × 10 = 20 dollars a month, after the doubling 40 dollars. The check costs 3 × 1500 + 3000 = 7,500 roubles once. Compare this with the cost of a forced urgent rework if you build your work on a model that does not suit you. If you cannot name that cost, estimate it before choosing a provider.

When this does not concern you

If you use a free chatbot and are not building your own work process on a model, the price per token does not matter to you. If you pay for the Plus tier or use a free version, then according to the article it is unclear whether you will get Argon and when.

Summary

On 13 October 2026 Google released Gemini 4 Argon for Pro and Ultra subscribers and said the model is on a par with the best in the industry. The claims of winning in tests were made by the company itself. The introductory price of 2 and 10 dollars per million tokens will double at the end of November. Before changing provider, test the model on 20 of your own tasks and count the cost of a task, not of a token.

See also another article in this section: Google gave an AI agent a work mailbox and a place in the company directory.

I work on AI agents and automation. If you want to look at my projects or discuss your task, visit my portfolio.

Sources