According to CNET, Google announced its new Gemini 3 Flash AI model on Wednesday, positioning it as a faster, lower-cost option that performs as well as previous powerful models. The model achieved a 90.4% score on the PhD-level GPQA Diamond benchmark and a 33.7% score on Humanity’s Last Exam. Google says it now outperforms the Gemini 2.5 Pro model from earlier this year at three times the speed. Gemini 3 Flash is available immediately for developers in Google AI Studio and is rolling out to general consumers in the Gemini app and AI Mode in Google Search. Enterprise customers can access it through Vertex AI and Gemini Enterprise.
The Speed vs. Power Trade-Off
Here’s the thing with these AI model announcements: there’s always a trade-off. Google is basically saying you can get near-top-tier reasoning without the computational cost and latency of its biggest models. That’s a huge deal for scaling real-time applications. Think customer service chatbots or in-game assistants—tasks where a two-second wait feels like an eternity. Gemini 3 Flash seems engineered for that specific niche: good enough, but blisteringly fast. The benchmark comparisons are telling. It’s close to Gemini 3 Pro (91.9% and 37.5% on those same tough tests) but not quite there. So the question becomes, for your use case, is that slight dip in absolute performance worth the massive gain in speed and cost? For many businesses, the answer will be a resounding yes.
Where You’ll Actually See It
Now, for the average person, this isn’t some abstract developer tool. It’s rolling out right into the Gemini app and, more importantly, the AI Mode in Google Search. That means when you ask Google a complex question, the summary you get back is increasingly likely to be powered by this faster, cheaper model. Google also mentioned it pulls “real-time local information and helpful links,” which is their not-so-subtle dig at competitors who can’t search the live web as effectively. But there‘s a catch for the fancier stuff. Want to use the even more powerful Gemini 3 Pro or the “nano banana” image model in Search? You’ll need an AI Pro or Ultra subscription. So the free tier gets the speedy workhorse, while subscribers get the thoroughbreds.
The Benchmark Game
Let’s talk about those benchmarks for a second. “PhD-level reasoning” sounds incredibly impressive, and scoring 90.4% on the GPQA Diamond test probably is. But I always get a little skeptical. These tests are designed to push models, but they’re also designed to be *benchmarked*. Companies optimize for them. Does scoring well on GPQA directly translate to giving a thoughtful, nuanced answer to a messy, real-world question? Not always. It’s a useful indicator, sure, but it’s not the whole story. The real test will be how it feels when millions of people start using it in Search for everything from homework help to planning a trip. Does it feel smart and fast, or is it just fast?
