Google is preparing the launch of a new artificial intelligence (AI) model with enhanced capabilities specifically for programming. Company employees indicate that Gemini 3.8 Flash, which uses the internal codename “Skimaki,” has the potential to narrow the competitive gap with rivals such as Anthropic and OpenAI, a sector of growing importance in the AI market.
Sources informed to The Wall Street Journal suggest that this model may be available as early as this Wednesday (2). During internal tests conducted on the programming tool called Jetski, some Google engineers demonstrated a preference for Gemini 3.8 Flash in direct comparison with the Opus model, developed by Anthropic.
Robust performance on industry evaluation benchmarks would be crucial for Google to address questions about its current standing in the race to develop AI models.
The significance of this situation was emphasized by recent changes in the company's AI division. Demis Hassabis, co-founder of Google DeepMind, left his executive leadership position last month as part of a restructuring. His replacement, Koray Kavukcuoglu, reportedly reinforced to employees the urgency of accelerating the pace of work.
Although Google gained prominence in the model race with the launch of Gemini 3.0 last November, the company faced subsequent challenges. The Gemini models fell behind the more advanced versions from Anthropic and OpenAI, especially in the field where AI agent-driven programming has established itself as a primary commercial use of the technology.
Additionally, the company lost some of its most notable researchers, including Noam Shazeer, co-founder of Character AI, and Jeff Dean, Google's chief scientist, both leaving the company earlier this year.
Regarding the larger Pro series models, modifications require a considerably greater allocation of Google's computational resources. This disparity allows the research teams at Google DeepMind and other laboratories to maintain multiple development strategies simultaneously, meaning that an unsatisfactory result in one specific model does not necessarily imply future problems for the company.
More information on the development
While focusing on the Flash versions, Google is behind schedule in the launch of a new model in the powerful Pro series. The company missed the planned timeline for a new version by several months, despite CEO Sundar Pichai assuring in May that the model would be available the following month.
According to people involved in the development, some candidate versions for Gemini 3.5 Pro were discarded because they did not show significant improvements over the Flash series.
The organization's next major model, Gemini 4, achieved good results in pre-training tests but still needs to complete the post-training phase.
Work on Gemini 3.7 Flash and Gemini 3.8 Flash began months ago, even before the management changes announced last month. According to DeepMind employees, Kavukcuoglu had been overseeing daily decisions regarding Gemini development since at least last year, while Hassabis dedicated much of his time to external commitments.
Previous reports by Business Insider already indicated that Google employees were testing Gemini 3.8 Flash.
Since the beginning of the year, Google has intensified the number of researchers and the computational resources dedicated to enhancing the programming capabilities of its models. One of the main bets lies in reinforcement learning, which is a stage after training where models acquire skills through trial and error.
The company has also strengthened its personnel in this area. Recently, Google hired Barret Zoph, former co-founder of Thinking Machines Lab and former head of OpenAI model post-training, to take on the role of Vice President of Research focused on reinforcement learning and post-training.
Current expectations are focused on the results of the benchmark tests and the performance of Gemini 3.8 Flash after its release.
