Google Deepmind is not backing away from the race to build frontier AI models. That is the message from Google Deepmind chief Koray Kavukcuoglu, who pushed back against the view that Google may be more focused on reach, cost, and price-to-performance than on leading at the top end of model capability.
His comments frame the company’s AI strategy around one central goal: staying at the frontier. At the same time, the details he shared were limited, leaving open questions about when Google’s next major models will arrive and how much progress has already been made.
A direct answer to doubts about Google’s AI priorities
Some observers have speculated that Google does not care as much about winning the AI model race as competitors do. The argument, as presented in the source, is that reach may depend more on price-to-performance than on holding the absolute lead in model capability.
Kavukcuoglu rejected that interpretation clearly. He said, "To put it very bluntly, there's nothing other than being at the frontier that is important for us. I'm 100% certain that we will be at the frontier."
That statement matters because it separates two ideas that are often discussed together in AI: broad usefulness and frontier performance. A model can be widely used because it is efficient, affordable, or well integrated. But Kavukcuoglu’s point is that Google Deepmind still sees frontier AI leadership as essential, not optional.
He also acknowledged that Google’s current models are "a little bit below the frontier." That admission gives the comments a sharper edge. Rather than claiming the gap does not exist, he argued that Google Deepmind has the team, resources, and full stack needed to close it.
Gemini 4 is described as ambitious, but without new details
The clearest forward-looking reference was Gemini 4. Kavukcuoglu described it as "the most ambitious run" so far and added, "touch wood, it's going well."
Still, the source makes clear that this was not a concrete product update. No specific frontier news was shared. The comments confirmed that Gemini 4 remains an important effort, but they did not provide a release date, benchmark, capability claim, or deployment plan.
That absence is important for readers trying to understand the state of Google Deepmind’s roadmap. The message is confident, but the public evidence described in the source remains limited. The company says it is aiming for the frontier, while the next major proof point has not yet been shown.
The source also notes that there was no update on Gemini 3.5 Pro. It says Gemini 3.5 Pro is apparently still in the works and is now months late. That leaves Gemini 4 and Gemini 3.5 Pro in different positions: one is being framed as an ambitious run, while the other remains expected but undetailed.
The Flash series points toward agentic coding
Instead of announcing a new frontier model, Kavukcuoglu highlighted progress in the Flash series. He described the move from 3.5 through 3.6 to 3.7 as a shift from a language model toward a coding agent.
His explanation was direct: "Basically, turn this whole thing into an agent, from a model to an agent," he said. He also added, "Software engineering is the most critical domain and environment that you want your systems to be successful at."
That emphasis shows where part of Google Deepmind’s model development is being tested: software engineering. In plain terms, the source presents the Flash series not only as a line of language models, but as a path toward systems that can act more like agents in coding workflows.
The distinction matters. A language model mainly produces responses. An agent, as described by Kavukcuoglu’s framing, suggests a system designed to take on more of a task-oriented role. The source does not provide implementation details, but it does show that Google Deepmind views coding as a critical proving ground for advanced AI systems.
The strategic message is confidence with unfinished evidence
The overall picture is a mix of strong ambition and incomplete public detail. Kavukcuoglu says Google Deepmind wants frontier AI leadership and believes it can get there. He also accepts that current models are slightly behind the frontier, while arguing that the organization has what it needs to close the gap.
For now, the practical signals are Gemini 4, the delayed Gemini 3.5 Pro, and the Flash series’ movement toward coding agents. Each points to a different part of the same strategy: build stronger frontier models, continue the Gemini roadmap, and make models more useful in software engineering environments.
What is missing is the concrete evidence that would settle the question. There was no new frontier announcement, no Gemini 3.5 Pro update, and no detailed claim about Gemini 4 beyond its ambition and current progress. Until those details arrive, the story remains one of stated priority rather than demonstrated outcome.
Still, the message from Google Deepmind’s chief is unambiguous. The company does not want to be seen as settling for reach alone. In Kavukcuoglu’s telling, price-to-performance may matter, but being at the frontier is the goal that matters most.