Gemini 4 Argon hit nearly 25T tokens in 72 hours, 3x Gemini 3

Google says its newly launched Gemini 4 Argon model processed nearly 25 trillion tokens in its first 72 hours, about three times what Gemini 3 handled in the same window after its own launch. The company shared the figure in a post on October 20.

In the post, Google wrote: “In the first 72 hours of our Gemini 4 Argon launch, we processed nearly 25T tokens, about 3x what we processed in the same time frame for Gemini 3.” It pointed readers to a conversation between CEO Sundar Pichai and Alex Kantrowitz on the Big Technology podcast. The post text we could read was cut off after the show’s name, and we could not access the interview itself.

What the number does and does not say

A token is the unit of text, code or other data that a model reads or writes. Counting tokens over a fixed window is a measure of how heavily a model is being used, whether by developers calling it through an API, by Google’s own products, or by agents that run long multi-step jobs.

The post gives only two data points: the total, and the multiple against Gemini 3. It does not say how the count is split between input and output tokens, between paying customers and internal Google traffic, or between consumer apps and developer tools. It also does not state the exact date Argon went live or the absolute Gemini 3 figure, though the multiple implies roughly 8 trillion or so for the earlier model. That back-of-envelope number is our arithmetic, not a Google statement.

The comparison is also a like-for-like launch window, not a comparison of two mature products. Token volume tends to rise with the size of the user base and with how much of the work is done by agents rather than chat, so a tripling could reflect a bigger Gemini audience as much as stronger demand for Argon specifically. Google has not published a breakdown, and the claim has not been independently verified.

Where 25 trillion sits next to Google’s overall traffic

For scale, Pichai said at Google’s I/O keynote on May 19, 2026 that the company was processing over 3.2 quadrillion tokens per month across its products, according to Shacknews. The same report lists his earlier figures: about 9.7 trillion tokens per month in 2024 and about 480 trillion in 2025.

Spread evenly, 3.2 quadrillion a month is roughly 107 trillion tokens a day. Argon’s launch volume works out to about 8 trillion a day, which would be on the order of 8% of that May daily average. That comparison is ours and is approximate: the two figures may be counted differently, and overall traffic has presumably kept growing since May. It does show that a single new model reaching this level within days is a meaningful slice of Google’s total load rather than a rounding error.

A usage claim, not a quality claim

Token counts say how much a model is used, not how well it performs. The post offers no benchmark results, latency figures, pricing or capacity data, and says nothing about whether the volume strained Google’s serving infrastructure. Readers weighing Argon for their own work still need the model’s documentation and independent tests for those answers.

The number fits a pattern in Google’s recent messaging. In the past week the company has promoted Argon through demos of agent-style work, including building a WebGL fur renderer inside its Antigravity coding environment and turning a photo of a desk into an interactive game with subagents. Those are the kinds of long, tool-using sessions that consume far more tokens than a single chat reply, which may help explain the speed at which the count climbed.

For developers and companies, the practical takeaway is limited to what Google has said so far: demand for Argon is high, and Google is willing to publicize it. Whether that translates into rate limits, pricing changes or capacity commitments is not addressed in the post.