Synopsis
Speaking to the media ahead of Google’s annual developer conference Google I/O, Pichai said the industry is entering a phase where efficiency and scale of deployment matter as much as raw model size.Listen to this article in summarized format
Speaking to the media ahead of Google’s annual developer conference Google I/O, Pichai said the industry is entering a phase where efficiency and scale of deployment matter as much as raw model size.
Citing Anthropic’s Claude Mythos, a model whose release has been restricted due to its ability to autonomously identify software vulnerabilities and generate working exploits, he noted that larger AI models still have a clear advantage in complex security tasks.
At the same time, the gap between massive frontier models and smaller systems is rapidly narrowing.
Google’s Gemini 3.5 Flash now delivers “almost 90% of the performance of frontier models” while being four times faster and available at roughly one-third to one-half the cost, Pichai noted. “Google’s Pro models are now reaching nearly 90% of Ultra-tier capability,” he added.
“What Mythos has done is show that there is value for the larger-size model in these kinds of security use cases. But it’s something we are capable of doing as well internally,” Pichai said.
He highlighted the scale at which AI adoption is accelerating across Google’s ecosystem. The company is now processing more than 3.2 quadrillion tokens per month across its platforms, up from 480 trillion tokens a year ago. Google’s APIs are also processing around 19 billion tokens per minute, while more than 8.5 million developers are building with Gemini models every month.
The company’s consumer AI products are also seeing rapid adoption. AI Overviews now has more than 2.5 billion monthly users, Pichai said, while AI Mode has crossed one billion users within a year of launch and the Gemini app has grown to more than 900 million active users now from 400 million last year.
Google’s focus is now centred on deploying AI agents at scale across products. “If you want to bring the power of agents and agent decoding to billions of users, how do you do that?” he said, describing affordability and efficiency as the central lens shaping Google’s AI roadmap.
Google also announced a wider push to embed AI agents across its ecosystem. The company introduced Gemini Spark, an AI assistant that can complete tasks across Gmail, Docs and other Workspace products in the background, alongside Gemini Omni, a multimodal model capable of generating video and other outputs from different forms of input. The company also unveiled Anti Gravity 2.0, a platform for building and managing autonomous AI agents.
Pichai also positioned Search as one of Google’s biggest AI advantages, calling it “the most used AI product in the world.” He said user behaviour has shifted sharply in recent years, with queries becoming longer, more conversational and increasingly multimodal across voice, image and video.
“The extraordinary thing about Search is how it keeps evolving to meet user expectations,” Pichai said. “Search has become less about individual queries and more like an ongoing conversation.”
Google also announced partnerships tied to AI transparency and commerce infrastructure.
OpenAI, Nvidia, Kakao and Lightricks are adopting SynthID, Google’s watermarking technology for AI-generated content, Pichai said.
Executives also said companies including Amazon, Meta, Microsoft, Salesforce and Stripe have joined Google’s Universal Commerce Protocol initiative aimed at enabling AI-powered shopping and payments.
On Google’s broader AI ambition, Pichai said, “It’s really about bringing all these frontier capabilities to billions of users around the world.”
