Gemini 4 Argon Processed Nearly 25 Trillion Tokens in Its First 72 Hours — Triple Gemini 3's Pace

Gemini 4 Argon's launch demand caught even Google off guard. CEO Sundar Pichai told the Sources podcast, hosted by Alex Heath, that the model processed almost 25 trillion tokens in its first 72 hours — roughly three times what Gemini 3.0 handled over the same period after its release. The surge has kept Google's capacity throttling in place, and Pichai acknowledged the company remains short on compute.

The interview also settled one of the more persistent mysteries of Google's model roadmap: what happened to Gemini 3.5 Pro, which was announced at I/O in May, never shipped, and was quietly written off by September. Pichai's answer is that the team killed it on purpose. While training the 3.5 series, researchers found real gaps in agentic coding. Rather than launch a Pro model with a known weakness, Google used its smaller, cheaper Flash models for fast iteration cycles.

That experiment worked better than expected. In roughly three months, the team shipped Gemini 3.6, 3.7 and 3.8 Flash in rapid succession, validating techniques for agentic and coding performance. By the time those lessons were ready to transfer to a Pro-tier model, Argon had already finished pre-training — and a 3.5 Pro built on the older base would have had almost no useful shelf life. Instead, Google injected the Flash-validated improvements directly into Argon's post-training.

Pichai also defused the rumored "Carbon" model, saying it is simply an internal codename in the frontier iteration process rather than a separate unreleased product. What comes next are distilled versions of Argon — smaller, cheaper models aimed at wider access — and he confirmed more Gemini 4 family models are on the way. Google has not said whether a Gemini 4 Pro will exist.

The most striking claim was organizational. Pichai said Google has entered what he described as a recursive self-improvement phase: models now directly participate in proposing training recipe suggestions and automatically evaluating the next checkpoint alongside researchers. He uses the workflow himself — an internal agent summarizes 24 hours of progress across pre-training, mid-training, post-training and data work daily, saving him about an hour of synchronization time, with telemetry-connected agents monitoring compute efficiency across teams.

He also addressed the DeepMind reshuffle that put Demis Hassabis into an Alphabet chief scientist role focused on safety, science and healthcare, with Koray Kavukcuoglu running day-to-day operations and Jeff Dean departing to found his own venture. Pichai framed it as a division of labor necessitated by scale: frontier model development now consumes so much data, compute, and coordination that its leadership structure has to specialize.

The interview's subtext is that Google now sees model releases as capacity problems as much as research ones. A 25-trillion-token debut week is a demand signal competitors would envy, but it also means every launch constrains the next one. Pichai's message to developers waiting for cheaper access: the distilled Argon variants are coming — built from the same lesson Google just learned in public, that in the current market the fast small models are how you earn the right to ship the big one.