Product Launch2 min read

The Price Per Token No Longer Tells What an AI Job Costs

By , Senior AI ConsultantPublished

Gemini 4 Neon matches Google's top model on a hard-task test for $3.74 a task against $7.76, because it uses far fewer tokens, so a model's price list says little about what a job costs.

Google released Gemini 4 Neon on November 18, and subscribers to Google's paid Pro and Ultra plans can already choose it in the Gemini app. It takes the place of Argon, which Google released on September 30 and opened to paying users in October. Developers can use it through Google's API, and free accounts do not get it.

The saving comes from fewer tokens, not only a lower rate

An AI model bills by the token, and the thinking it does before answering is billed too. So what a job costs is the price of a token times the number of tokens the model uses to finish it. Neon's price is lower than Carbon's standard rate: $3 against $4 per million tokens going in, and $15 against $20 coming out.

On a hard-task test, Neon at its high setting scores 60, the same as Carbon, for $3.74 a task against $7.76. A rate that is 25% lower takes $7.76 down to about $5.80. The rest of the way to $3.74 has to come from Neon using roughly a third fewer tokens to reach the same score.

The setting is a second dial. Neon's top level scores 61 for $6.44 a task, so that one extra point costs about 70% more than the high setting, which for most work is not worth paying.

For example, a small finance team that spends $2,000 a month having Carbon read and check supplier documents at its standard price would pay about $960 for Neon at high, if its documents behave like the test tasks.

Carbon's discount ends on November 30 and changes the comparison

Carbon and Argon cost $2 and $10 per million tokens through November 30, and $4 and $20 from December 1. Against the discounted rate Neon is 50% dearer per token, which is why it looks expensive in November. Carbon's discounted cost for the same test task works out to about $3.88, which is a tie with Neon's $3.74.

On December 1 Carbon's bill doubles and Neon's does not. A tool built on Carbon in November will cost twice as much in December unless someone moves it. We showed with Argon that a launch price hides what tasks cost, and Neon is a second case: the price on the list and the bill at the end of the month only line up when two models use the same number of tokens.

In the app, the choice is now by kind of work

If you pay for Gemini rather than build on it, you never see a token, but the same arithmetic reaches you through speed and limits. The model picker now names kinds of work: Everyday help runs on Hydrogen, Deeper analysis on Neon, and Code and complex work on Carbon, while Auto chooses for you. Google's own examples for Neon are household ones, such as finding out why a utility bill jumped, or checking several receipts against each other to win a refund.

Neon also writes faster, at about 110 tokens a second against about 57 for Carbon. At its hardest setting it finishes a test task in about 21 minutes, where Carbon at high takes about 36. And because it needs fewer tokens for the same result, a plan's usage limit should last longer.

Google has now released four Gemini 4 models since September 30, and the picker will change again. The comparison that survives each change is your own: run twenty real jobs from your work on two models, write down what each costs and how long it takes, and keep the one that finishes cheapest. Do it again when the next model arrives.

Share this

STAY INFORMED

Get AI intelligence like this delivered to your inbox.

Free forever · Unsubscribe anytime


You May Also Find Valuable