Google Gemini 3.8 Flash coding AI model expected this week

Google's AI research unit plans to ship a new model boasting sharply improved coding abilities as soon as Wednesday, the Wall Street Journal reported. The model, Gemini 3.8 Flash, is referred to internally as "Skimaki" and is designed to close ground on rivals Anthropic and OpenAI in an area where Google has lagged.

Engineers at Google have tested Skimaki against Anthropic's Opus model using Jetski, Google's internal coding tool, and have preferred the new model, according to the Journal. The release comes after Google has devoted more resources to reinforcement learning — a later stage of model training that teaches models to perform skills through trial and error — since the start of the year.

A strong showing by Gemini 3.8 Flash would not by itself reestablish Google's position at the frontier of AI development. The Flash series is built to be smaller, cheaper, and faster to run than Google's flagship models, which are built from trillions of numerical parameters.

The release follows leadership changes at Google DeepMind. Demis Hassabis, the unit's co-founder and Nobel Prize winner in chemistry, stepped aside as chief executive last month. His successor, Koray Kavukcuoglu, has told employees he wants to increase the pace of execution. Kavukcuoglu had already taken charge of day-to-day Gemini decisions for at least a year before his appointment, while Hassabis directed his attention toward outside commitments.

Google has also struggled with its larger model lineup. Internal candidates for Gemini 3.5 Pro were scrapped because they did not represent a sufficient improvement over the Flash series. Chief Executive Sundar Pichai had said in May that a Pro model would arrive "next month," but none has materialized. Gemini 4, Google's next planned flagship, posted encouraging numbers in pretraining evaluations but has yet to finish the posttraining phase.

Google introduced Gemini 3.7 Flash roughly three weeks before this latest release, marketing it to businesses developing autonomous AI systems as a lower-cost option. That model showed gains on coding benchmarks including FrontierCode 1.1 and DeepSWE v1.1 compared with its predecessor.

Google also brought on Barret Zoph, who previously co-founded Thinking Machines Lab and served as OpenAI's posttraining lead, to fill a vice president of research role covering reinforcement learning and posttraining.