DeepMind is adopting a new methodology for deploying intelligence models that emphasizes continuous iteration over waiting for total maturity. Google will launch an early post-training version of its next flagship AI, Gemini 4, as soon as possible to allow for rapid updates, according to chief Koray Kavukcuoglu. The model is currently in the refinement stage and will arrive much earlier than the end of the year.

The strategy intends to close the performance gap with rivals OpenAI and Anthropic after nearly a year without a flagship release. Gemini 4 is active internally through a system called Antigravity and is informing the design of the next 2 or 3 generations of Google's Tensor Processing Units (TPUs). To increase development speed, Google is shifting focus toward Flash models and moving past Gemini 3.5 Pro while it begins selling TPUs directly to compete with Nvidia.