Google is about to make a key product update in the field of artificial intelligence. According to insiders, Google plans to officially launch a new model called Gemini 3.8 Flash, which has significantly improved programming capabilities, as early as Wednesday local time. The internal development codename for this model is "Skimaki."

In internal Jet Ski programming tool comparison tests, engineers showed a stronger preference for the new model than Anthropic's Opus model. Internal test data shows that Gemini 3.8 Flash has significantly narrowed the gap in core programming capabilities between Google and the flagship models of Anthropic and OpenAI.

image.png

This new model's release comes at a time when the Google DeepMind team is undergoing significant personnel changes. Last month, DeepMind co-founder Demis Hassabis stepped down from daily management duties and took on the roles of department chairman and Alphabet Chief Scientist. His long-time deputy Koray Kavukcuoglu has fully taken over operations and was promoted to Senior Vice President, and since taking office, he has clearly required the team to significantly accelerate execution speed.

Looking back at recent periods, Google faced many setbacks in the intense model competition. Since the release of Gemini 3.0 last November, its product line gradually lagged behind competitors in key performance aspects such as programming. In addition, several core researchers, including Noam Shazeer, co-founder of Character.AI, and Chief Scientist Jeff Dean, left the company during the summer of this year.

From a product positioning perspective, Gemini 3.8 Flash still belongs to the Flash series, which is smaller in size, faster in operation, and lower in cost. Its overall performance does not equate to the Pro series flagship model that includes trillions of parameters. It is reported that the release of the higher-positioned "Pro" series new model has been delayed for several months. The previous 3.5Pro candidate model was abandoned because the improvement compared to the Flash series was not significant enough. However, the next-generation flagship model Gemini 4 performed well in pre-training evaluations, but the post-training work is not yet completed.

To continuously strengthen the underlying capabilities of AI models, Google has shifted more R&D time and computing power resources to the reinforcement learning aspect this year. In addition, the company recently successfully recruited Barret Zoph, co-founder of the former Thinking Machines Lab and former head of pre-training at OpenAI, as Research Vice President, who will be fully responsible for technical breakthroughs in reinforcement learning and post-training areas.