DeepMind’s New Leader Signals Gemini 4’s Imminent Arrival, Aiming to End Google’s AI Flagship Drought
Google DeepMind’s newly appointed head, Koray Kavukcuoglu, has broken his silence in his first public interview, revealing that Gemini 4 has entered its post-training phase and is slated for an early release, potentially before the end of 2026. This move aims to fill a significant gap in Google’s flagship AI model offerings, which has persisted for nearly ten months.
A Ten-Month Flagship Hiatus
For the past ten months, Google has not introduced a new flagship AI model. During this period, competitors have surged ahead: OpenAI has unveiled GPT-6, and Anthropic has launched its Mythos series. Both companies’ new models reportedly surpass the performance of Google’s Gemini 3, released in November 2025.
Now, Koray Kavukcuoglu, who has taken the helm at Google DeepMind, has addressed the situation. Speaking at The Information’s AI Agenda Live summit, he confirmed that Gemini 4 is in its post-training phase, a crucial stage involving fine-tuning and human feedback correction after the primary training is complete.
Kavukcuoglu disclosed that the new model is undergoing internal testing on Google’s proprietary development platform, Antigravity. The ambitious target is to launch Gemini 4 “significantly ahead” of its initially projected end-of-2026 release.
Leadership Shuffle and Strategic Pivot
This period of flagship model absence coincided with a significant reshuffling of Google’s AI leadership. On August 5, 2026, Sundar Pichai and Demis Hassabis announced in a joint Google blog post that Hassabis would step down from day-to-day operations at DeepMind. He transitioned to the role of Chairman of Google DeepMind and Chief Scientist at Alphabet, focusing on Artificial General Intelligence (AGI) and Alphabet’s AI drug discovery company, Isomorphic Labs. Koray Kavukcuoglu, a 13-year veteran of DeepMind, was promoted to Senior Vice President, reporting directly to Pichai and overseeing Gemini model development, cutting-edge AI research, and the Gemini App team.
In his blog post at the time, Hassabis expressed confidence: “I have always been incredibly confident in handing over the Gemini models to Koray and the team, and I am excited about the progress we are making on new models like Gemini 4.”
Google’s previous flagship, Gemini 3.5 Pro, has faced multiple delays. At the I/O conference in May 2026, Pichai publicly pledged that an enhanced version of Gemini 3.5 Pro would be released in June to close the gap with competitors. However, the June, July, and August targets were missed, and the model has yet to be released.
Kavukcuoglu explained this delay by stating the team “took a small step back” to reallocate resources to the faster, smaller-scale Flash models. At the time, the team determined that prioritizing rapid learning speed was more advantageous than pushing a flagship model that was proving difficult to perfect.
This strategic choice, Kavukcuoglu emphasized in his recent interview, underpins Gemini 4’s strategy: to expedite the release of early post-training results. “We’re seeing the results, and we’re excited,” he stated, adding that the team plans to continue with “rapid iteration.”
Redefining Success Beyond AGI
When questioned about Google’s perceived lag and whether it would follow Hassabis’s pursuit of AGI, Kavukcuoglu deliberately sidestepped the framing. “Whether we achieve AGI is not the point of the discussion; the real question is whether we can build trustworthy intelligent agents,” he remarked. Addressing concerns about falling behind rivals, he expressed “100 percent confidence,” asserting, “In my view, we are always at the forefront; that’s a certainty.”
Despite the delays in its flagship model, Google’s current business metrics are robust. The Antigravity platform boasts over 2.4 million weekly active users, and its model APIs process approximately 22 billion tokens per minute. While the user base remains strong, the critical missing piece is a flagship model capable of directly competing with GPT-6 and Mythos.
However, as with any technological promise, until the model is officially launched, it remains an unfulfilled commitment.
