What Google has advanced about Gemini 4

Last week, Google teased Gemini 4 quite a bit, as 3.5Pro it missed its launch in June.

Gemini 4

The first mention of the, unsurprisingly, so-called “Gemini 4” coincided with the announcement of Gemini 3.6 Flash3.5 Flash-Lite and 3.5 Flash Cyber. While Google is testing the Gemini 3.5 Pro with partners ahead of its launch “as soon as it’s ready,” work is underway on the “next generation of models.” Specifically, this is the company’s “most ambitious pre-training execution yet.”

To the Alphabet Q2 2026 Earnings CallSundar Pichai reiterated that Google is “excited about the progress we are seeing at the border.”

During the Q&A session with analysts, the CEO was asked about his “level of confidence that the Gemini models can remain on the frontier.”

Announcement: Scroll to see more content

Look, I think the border is an incredibly dynamic space and it’s advancing ferociously. At any given moment, when you take a snapshot, it feels dynamic. We have had clearly frontier models. There are many attributes where we are still on the frontier. There are areas where we have recognized that we need to improve. Coding and agentic coding are an example of that, and I think the teams are very, very focused on it.

Pichai explicitly said that Google is “very committed and very confident about being on the frontier.” Additionally: “We want to compete at the frontier level where the frontier will be when Gemini 4 comes out, so we’re applying a lot of our calculation and effort in that direction.”

In another question, Pichai was asked how Google is balancing internal and external customer demand for tensor processing units (TPUs):

In terms of TPU allocation, look, our first priority is to make sure that we are allocating what we need to compete on the frontier in terms of AGI development. And that is the basis of everything we do.

The closest we got to details this week was that Gemini 4 will need a “larger base model to compete on that frontier model.”

For the next generation of Frontier, you’ll need much larger base models. We are now training Gemini 4 and are being very ambitious with it. I’m very excited about the progress I’m seeing internally on Gemini 4 and I’m sure people will be happy when we take it out.

According to previous releases, Gemini 4 should launch in November or December.

Gemini Flash

Pichai also said he expects continued iteration on Flash Gemini 3.x models “with further advancements in agent coding, etc.” Specifically, Google aims to “release models at almost a monthly cadence.”