Google pushes Gemini 4 as DeepMind leaves AGI behind

Gemini 4 is being pushed toward release

Kavukcuoglu made his first appearance before journalists as head of Google DeepMind at The Information AI Agenda Live Summit. He said Gemini 4 is in the early post-training stage, where a base model is refined to perform tasks more reliably.

Google wants to show the first post-training result as soon as possible because the company is already seeing capabilities it likes. Kavukcuoglu expects rapid improvement cycles to follow.

Safety work is still underway. Engineers are already using the model internally in Antigravity, Google’s programming tool.

Kavukcuoglu previously served as chief technology officer and replaced Demis Hassabis as head of DeepMind in August. Hassabis is now chairman of the division and remains Alphabet’s chief scientist.

The release schedule also reflects unfinished business:

Kavukcuoglu acknowledged that Google had to take a small step back. He said the company had prioritized accelerating training as much as possible, but did not clarify whether Gemini 3.5 Pro had been officially canceled.

Google is trying to close the gap

The competitive gap has widened while Gemini 3.5 Pro remained absent:

Kavukcuoglu said he was not worried about Google’s position. He expressed complete confidence in himself and in a team he considers exceptionally strong, arguing that Google will inevitably remain among the leaders.

The model race also feeds Google’s chip business. This year, the company began selling its tensor processing units directly to customers, intensifying competition with Nvidia. Kavukcuoglu said new models show hardware developers how to design the next two or three generations of chips.

Google has caught up before. After Gemini 3 launched, OpenAI declared “Code Red” and only reached parity with Google after GPT-5.2. Whether Gemini 4 can repeat that result is still unknown.

My read is that the urgency around Gemini 4 is partly defensive. Google is not announcing a clean new lead; it is trying to make its next flagship arrive before the distance becomes harder to explain.

The AGI mission is receding

Kavukcuoglu’s comments about AGI are as revealing as his comments about timing. He said asking whether the company has achieved AGI is no longer the right question. The more important issue, in his view, is whether developers can build reliable AI agents that people can trust.

That is a sharp change from DeepMind’s earlier identity. Only weeks before Kavukcuoglu took over, Hassabis wrote in a farewell letter that he had spent his life working toward AGI and, like many others, felt that it was approaching.

Hassabis founded DeepMind with other researchers in 2010. Google acquired the lab four years later, and the London team retained an unusual status inside the company for a long time. ChatGPT changed that arrangement. During Google’s first “Code Red,” Google Brain and DeepMind were combined into Google DeepMind, where the Gemini models were also named.

The August restructuring completed much of that shift:

How voluntary Hassabis’s departure was remains unclear. The Pathfounders newsletter, citing industry sources, reported that he wanted to leave at the same time as Jeff Dean, but that Google leadership feared a collapse in the stock price. Google did not comment on those reports.

The loss of researchers is less ambiguous:

According to Fortune, the ratio of new hires to departures changed from about 12 to 1 in the second quarter of 2023 to about 2 to 1 in the third quarter of 2026.

A research lab becomes a product division

This transition comes as the argument over AGI and frontier-model risks is intensifying again. On September 12, Anthropic CEO Dario Amodei published an essay titled “We Must Pace the Frontier.” He called on frontier-model labs to slow the rate at which they increase model capabilities, without stopping development entirely. OpenAI supported the appeal, Elon Musk agreed with it, and Hassabis called its direction correct while saying the details needed more work. In July, more than 1,200 employees at labs building frontier models signed a letter with the same title.

Kavukcuoglu is not dismissing safety. Reliable AI agents are themselves a safety concern. But while his predecessor and Google’s competitors argue over AGI and its risks, the new DeepMind chief is framing model development mainly as a product problem.

I think that distinction is the real story behind Gemini 4. The question is no longer only whether Google can produce a stronger model. It is whether a division created to pursue AGI can still preserve that research identity while every organizational decision pushes Gemini toward commercial delivery.

The pressure began with ChatGPT and has continued through OpenAI, with Anthropic now adding to it. DeepMind has become Google’s Gemini factory. With its leadership, teams and development center reorganized around that role, the transition looks complete.