Gemini 3.5 Flash: Fast Enough for Gen AI to Make Sense
▼ Summary
– Google is rolling out Gemini 3.5 Flash across its products, claiming it surpasses the previous Pro model in performance.
– The model offers frontier-level intelligence while being efficient enough to make complex agentic tasks viable at scale.
– Gemini 3.5 Flash outputs nearly 300 tokens per second, with benchmark scores similar to larger frontier models that are four times slower.
– Improvements in pre-training and post-training feedback from developers, such as from Antigravity, have enhanced code and tool use performance.
– Google plans to continue this trend, with future releases like 3.5 Pro expected to be better, and subsequent Flash models matching Pro performance.
At last year’s I/O, Google was still showcasing the Gemini 2.5 series. A single year later, the pace of change is remarkable. The company has since moved through the 3.0 and 3.1 families, and now Gemini 3.5 Flash is here. Rolling out across a wide array of Google products starting today, this latest model once again claims to outperform its predecessor’s Pro version.
This pattern of iterative improvement has defined Google’s recent release cycle, but the team insists this update is different. Gemini 3.5 Flash is said to deliver frontier-level intelligence while being efficient enough to make complex, agentic AI tasks economically viable at scale. Tulsee Doshi, Google’s senior director of product management for Gemini, notes that the model’s breakthroughs are woven into multiple products, and this launch is only the beginning.
Generative AI remains a costly endeavor, and every major player is searching for efficiency gains. The challenge intensifies when building agentic experiences designed to run longer and handle intricate workflows. Gemini 3.5 Flash could be a pivotal step toward making those use cases practical. The model processes nearly 300 tokens per second, yet its benchmark scores rival those of larger frontier models like 3.1 Pro, which generate outputs at roughly a quarter of that speed.
Doshi explains that while pre-training improvements were significant, the real breakthroughs came from post-training insights gathered from developer usage. “With post-training, we’re really starting to unlock some of the value of the feedback we’re getting from users, for example, from Antigravity,” she said. “That’s really what you’re seeing play out in terms of the code performance and the tool use performance. And then, the hope is that you’ll continue to see the step change where 3.5 Pro will be better, and the next Flash meets Pro performance with that series.”
(Source: Ars Technica)
