Google's new Gemini 3 Flash AI model is built for faster output at a lower cost, and functions as well as previous powerful reasoning models, Google said in a press release on Wednesday.
According to benchmark tests, Gemini 3 Flash achieves PhD-level reasoning in the GPQA Diamond test, with a score of 90.4%, and achieved a score of 33.7% (without tools) in Humanity's Last Exam. These are very difficult tests designed to push AI models and require expert-level knowledge. By comparison, Gemini 3 Pro achieved scores of 91.9% and 37.5%, respectively. Google said Gemini 3 Flash outperforms Gemini 2.5 Pro, which was Google's top model when it released earlier this year, at three times the speed.
Don't miss any of our unbiased tech content and lab-based reviews. Add CNET as a preferred Google source.
Gemini 3 Flash is available in Google AI Studio and Gemini CLI for developers. For general consumers, it's rolling out in the Gemini app, the new Antigravity and AI Mode in Google Search. For enterprise users, Gemini 3 Flash is available in Vertex AI and Gemini Enterprise.
Google said Gemini 3 Flash can be a handy agent for customer support or in-game support, tasks that require fast responses.
"This is a model that is smart and fast," said Josh Woodward, vice president of Google Labs and Google Gemini in an interview with CNET. "A lot of times you have to make a trade off. You go for a big model that's slow or a fast model that's dumb. And this one is right at that point on the Pareto frontier, when we look at quality and speed, that is a very interesting spot."
Think of the Pareto front as finding a solution with the best possible compromises. For example, when trying to buy a car that's both fast and cheap, there will need to be some trade-offs in body material or seat quality. The Pareto front in engineering looks to find the most efficient solution with the fewest consequences.
Woodward sees Gemini 3 Flash as a model to tackle everyday tasks, such as questions about travel, shopping or education. Interestingly, Google will also introduce a "thinking" version for Gemini 3 Flash, which will take a bit more time to formulate responses. It's the first time Google is doing this and Woodward is curious to see how consumers will react. Google is also experimenting with an "auto" mode, one in which the model intuitively switches between fast and thinking versions based on the query.
Gemini 3 Pro and nano banana, Google's image model, are landing in AI Mode in Search. They'll be accessible via a dropdown menu and is only available to AI Pro and Ultra subscribers. For those who want to stick to the free tier, Gemini 3 Flash will be in AI Mode and users can choose the thinking model as well for better, if slightly slower, output.
Gemini 3 Flash comes as Google is starting to gain a lead against ChatGPT maker OpenAI. Last month's release of Gemini 3 reportedly pushed OpenAI into declaring a "code red." OpenAI has since fought back with last week's release of GPT-5.2. Anthropic also released a new version of its most powerful model, Claude Opus 4.5, in recent weeks.