Google Confirms the Upcoming Launch of the Next-Generation Flagship Gemini 4: Has Entered Post-Training and May Be Released Much Earlier Than the End of the Year

At The Information's AI Agenda Live event on the 23rd local time, Demis Hassabis, head of Google DeepMind, revealed that Google's new flagship model Gemini 4 is about to be launched and has entered the "early post-training" phase of the development process. Post-training refers to a stage where the base model is further adjusted and its behavior reliability is improved.

Skip Gemini 3.5 Pro and Focus on Gemini 4

Hassabis hopes Gemini 4 will be released "well before the end of the year" and said they are quickly delivering early results from the post-training phase, "because we saw the results and were very excited." After that, Google will continue to rapidly iterate. Gemini 4 will be Google's first next-generation flagship model since the Gemini 3 series at the end of last year. In February this year, Google released an upgraded version of Gemini 3.1, and later launched several smaller-scale, lower-cost but less complex reasoning Flash models.

Looking back, in May this year, Google CEO Sundar Pichai had announced at the I/O conference that a major update to Gemini 3.5 Pro would be released in June, but it ultimately did not appear. Hassabis explained that after the release of Gemini 3 and 3.1, Google did not continue with the 3.5 Pro but instead "took a step back," focusing on Flash models, "for us, the most important thing at that time was to accelerate learning as much as possible." He did not clarify whether 3.5 Pro was officially canceled, but clearly stated that the current focus is on launching Gemini 4.

Already Used in Internal Programming Tools, Safety Testing Underway

Asked if he was worried that Google was falling behind competitors, Hassabis responded that he had "100% confidence in the team," "our team is incredibly strong, and I have the greatest trust in them; we will always be at the forefront of technology." Currently, Gemini 4 is in the stage of developing protective mechanisms and safety testing, and Google engineers have already used it to run the Antigravity AI programming tool internally. Unlike some competitors, he did not express a clear stance on whether the industry has achieved AGI, only stating, "discussing whether AGI has been achieved is not the right direction; the real issue is whether we can build trustworthy agents."