Google CEO Sundar Pichai acknowledged that his company is lagging behind rivals in agent coding, one of the industry’s most lucrative AI capabilities. speak at hard fork In a podcast published by the New York Times, Pichai said that while Google’s models are at the forefront of text, multimodality, voice, and inference, that’s not the case when it comes to long-term coding tasks, where Anthropic’s Claude Code and OpenAI’s Codex have become calling cards. “I think we’re a little behind the curve at the moment when it comes to agent coding, which involves using tools, following instructions, and long-term tasks,” he said.His explanation was straightforward and concerned the product, not the quality of the model. Google didn’t have a physical place for developers to work, so no data was leaked back. “If you use Claude Code as an example, we may not have fully scratched the surface yet,” Pichai said, adding, “I’m very, very optimistic and confident that we’ll get through there.” The entrance ends on an awkward note. Pichai argues that while Google is the only large company truly at the frontier, that frontier is defined by two startups that took one problem and refused to let it go.
Why agent coding has become the most valuable battleground in the AI industry
Money was spent on coding. Anthropic’s revenue exploded thanks to Claude Code, OpenAI pivoted from consumer to enterprise with Codex, and one executive reportedly told staff to stop chasing side quests. Citing a Microsoft study that found that using AI assistants made developers work more than 50% faster, Wired pointed to the shape of this battle in 2023, arguing that the real battle was not about chatbot parlor tricks, but about which companies could attract developers first. Mordor Intelligence currently expects the AI code tools market to grow from $9.3 billion this year to approximately $30 billion by 2031.Pichai rejected suggestions that Google’s widespread betting is the problem. “I don’t think it’s a matter of focus,” he said. hard fork. “We are a large company and our scale allows us to focus on multiple things at the same time.” He described the tempo on the ground in words that sounded more like a reassurance than a warning: “30 to 60 days seems like five years.”
Antigravity 2.0 and Gemini 3.5 Flash: Google’s answer to Claude Code
Google’s fix was published in I/O 2026. Antigravity 2.0 shipped as a standalone desktop app with a CLI, SDK, and the ability to coordinate multiple agents at once. Gemini 3.5 Flash, co-developed with Antigravity, is claimed to be four times faster than competing Frontier models, with Google claiming that the optimized version is up to 12 times faster. On stage, engineer Varun Mohan showed how agents emerge to build individual components of an operating system from scratch, a process Pichai said takes thousands of human hours.He also acknowledged a difficult start. The tightening of usage limits prompted complaints from developers, and Google subsequently reset quotas. “When you come across this, it’s understandable that it’s a source of frustration,” he says. “I feel the same way.”
Google’s AI coding gap: What the internal numbers really show
Pichai’s evidence that the gap is closing is in internal hiring. Token usage within Google is doubling every week, he said, but “we’ve never seen anything like it” within the company. By using anti-gravity in the real world and pulling back that data, he hopes to make a big difference.However, internal numbers go both ways. Google’s own CFO said that Anthropic writes almost 100 percent of its code with AI assistance, while Google only writes about 50 percent. The AI Coding Strike team, formed in April with Sergey Brin and DeepMind CTO Koray Kavukcuoglu, is already being restructured. Brin wrote in an internal memo that Google needs to “urgently close the agent execution gap” and transform its models into the primary developer of final code.Another lever is price. Google introduced a $100 per month AI Ultra tier on I/O and lowered the price of its top plan from $250 to $200, positioning itself as a cheaper option for heavy coding workloads. It only works if the tools can withstand it.
