By Erin Woo

Google's artificial-intelligence research unit is set to release a new model with significantly upgraded coding capabilities, according to employees, potentially shoring up an area where the company has lagged behind competitors OpenAI and Anthropic.

The model, dubbed 3.8 Flash and referred to inside the company as "Skimaki," is scheduled for release as soon as Wednesday, the people said. In head-to-head testing within Jetski, Google's internal coding tool, the company's engineers have preferred it to Anthropic's Opus model, some of the people said.

Strong performance on industry-standard benchmark tests could answer some of the questions that have surrounded Google DeepMind since before its co-founder, Demis Hassabis, stepped aside in an executive shake-up last month. Hassabis' successor, Koray Kavukcuoglu, has emphasized to employees that he wants to increase the pace of execution.

After a brief turn as a front-runner in the model-development race with the release of Gemini 3.0 last November, Google has encountered a series of setbacks this year. Gemini has returned to trailing flagship models from Anthropic and OpenAI at a moment when agentic coding has taken off as the premier business use case for AI. Some of Google's highest-profile researchers, including Character AI co-founder Noam Shazeer and Chief Scientist Jeff Dean, left earlier this summer.

A strong debut by Flash 3.8 wouldn't by itself likely be seen as re-establishing Google's place on the so-called frontier. The Flash model series is designed to be smaller, cheaper and faster to run, but lacks the power of the biggest models, which consist of trillions of numerical parameters.

Google has fallen months behind schedule in releasing a new model from its most powerful "Pro" series, despite Chief Executive Sundar Pichai saying in May that one would arrive "next month." Internal candidates for 3.5 Pro were scrapped because they weren't sufficiently better than the Flash series, the people said. Gemini 4, the company's next planned flagship model, has performed well on pretraining evaluations but still needs to complete posttraining, according to the people.

The small size of Flash models means they require less computing capacity to alter, allowing multiple teams to try approaches out and develop them at once, the people said. In contrast, making changes to Google's bigger Pro models requires corralling much more of Google's resources.

Inside Google DeepMind and other labs, research teams constantly have multiple bets going at once, meaning that lagging performance in one model under development doesn't necessarily have a bearing on other coming models, the people said.

Google's recent 3.7 Flash and 3.8 Flash models were in the works for months, with their development beginning well before last month's leadership changes, some of the people said. Kavukcuoglu has been leading day-to-day decisions on Gemini development since at least last year, with Hassabis devoting much of his time to external engagements, according to the DeepMind employees.

Business Insider previously reported that Google employees had been testing 3.8 Flash.

Since the beginning of the year, Google has been pouring more of its researchers' time and computing resources into improving its models' coding capabilities. In particular, it has increased the resources devoted to reinforcement learning, a step later in the training process that teaches models to perform skills through trial and error, the people said.

The company recently hired Barret Zoph, a former Thinking Machines Lab co-founder and OpenAI posttraining lead, to be a vice president of research focused on reinforcement learning and posttraining.

Write to Erin Woo at erin.woo@wsj.com

(END) Dow Jones Newswires

September 01, 2026 17:06 ET (21:06 GMT)