It’s one of the more open areas of competition in AI right now. In an unvarnished talk with The New York Times, Pichai put it plainly: on agentic coding – the kind of AI that can be put to work on software tasks for the long haul – we are behind the likes of Anthropic and OpenAI. What to do? He says to put your head down, ship and let the users teach you.
The competitive gap in agentic coding
Pichai won’t have it any other way than to say we are trailing on the tool use and instruction following side of things. The competition got there first with the sort of involved, long-running code that a developer in the field would want to depend on.
But don’t read this as a step back. It’s a matter of priority. You can’t get around the fact that coding is at the heart of what we do, so we are making a point of closing the distance.
Why rivals pulled ahead
If you ask Pichai, it wasn’t a case of inferior models. It was about the data. Other companies had their developer-facing products scale sooner, which in turn built better feedback loops.
Look at something like Claude Code or Cursor. That’s where the developers were, and that gave our competitors a running start in this space.
Google’s catch-up playbook
For now, we have Gemini 3.5 Flash. Pichai is touting it as the fix for the soft spots in our coding-heavy work, and he is on about the value of getting in front of real-world data to make improvements.
We’re also putting our own advanced system, Antigravity 2.0, to the test in-house. Pichai says we’re seeing usage go up two-fold week in and week out. The numbers from I/O on token consumption make it clear our people are using these models in a big way.
In short, here is where Pichai stands:
– We are not where we should be on agentic coding
– But our models are still top-tier for general intelligence
– Gemini 3.5 Flash is meant to shore up our coding side
– And internally, we are doubling down every week
Strength in general AI remains intact
Pichai made sure to stress that when it comes to pure intelligence, we are at the very edge. Whether it’s text, voice, audio or just plain reasoning, we are in a leading position.
There has been a lull in some of the more niche programming tools, he concedes. But with every lab on its own pretraining schedule, the pace of new things coming out is all over the map. It’s a moving target.
What it means for developers and the market
You can take it from Pichai at face value. If Gemini 3.5 can handle the harder, longer agents and we can mine the data from it, the whole thing will pick up speed. That is the only way to even the score with Anthropic and the rest on the kinds of coding you do every day.
He is having none of the talk that we are out of it. We know where the line is, we are moving fast with our own deployment, and we’ll be fine.
The stakes in the coding-assistant wars
This is becoming a make-or-break area. It’s about who the developers stick with and how quickly an enterprise can update its software. Admitting we were late to the data is one thing; using Gemini 3.5 and our own heavy internal traffic to get back on top is the answer.
Coding is what we build on, Pichai will tell you. The next round of training will tell us if we have the ground to stand on to make a quick return to the front of the pack.
