What “$2 per million tokens” really means (using Gemini 4 Argon’s price)

What a token is

Language models read and write text in small pieces called tokens. A token is often part of a word, so a short word may be one token and a long or unusual word may be several. As a rough rule of thumb, a page of ordinary English is a few hundred tokens, though exact counts vary by model.

Pricing for developers is quoted per million tokens because single requests are tiny. Input tokens are what you send, such as your question and any documents you paste in. Output tokens are what the model writes back. The two usually have different prices, and output is usually the more expensive of the two.

The Gemini 4 Argon numbers

Coverage of Google’s launch on October 13, 2026 reports an introductory price of $2 per million input tokens and $10 per million output tokens on the paid Gemini API, with a large discount on cached input. Cached input means text the model has already been given and can reuse.

The reports do not agree about what comes next. One roundup says the rates rise to $4 and $20 after December 1, while another says the later rate is not stated. If you are budgeting, check Google’s own pricing page rather than relying on either figure.

A worked example

Suppose you send a request with 2,000 input tokens and the model replies with 500 output tokens. At the reported introductory rates, the input costs 2,000 divided by 1,000,000, times $2, which is $0.004. The output costs 500 divided by 1,000,000, times $10, which is $0.005. The request costs about nine-tenths of a cent.

That looks cheap, and for one request it is. The numbers add up when an app makes thousands of requests a day, or when each request carries long documents. An agent that makes many calls to finish one task can use far more tokens than a single chat message.

Why the headline price can mislead

• Introductory prices can end, and reports disagree on what replaces this one.
• Output tokens cost more than input tokens, so a chatty model can cost more than its input price suggests.
• Long prompts and long replies both count. Pasting in a big document on every request is the usual cause of surprise bills.
• Caching can lower costs a lot, but only if your app is built to reuse the same text.
• Company-reported performance claims do not tell you what a task costs. Google’s benchmark claims were not independently verified in the coverage we read.

What this means if you are not a developer

If you use the Gemini app, you pay through a subscription rather than per token, and the coverage we read names Google AI Pro and Ultra subscribers for the new model. The token price matters if you or your company build on the API. Otherwise, treat it as a signal of how vendors compete: lower prices and more caching tend to follow each new release.

A sensible habit is to test a model on three real tasks, count the tokens used, and compare the total cost against the quality you got. We have not tested Gemini 4 Argon ourselves.