Koray Kavukcuoglu, the new head of Google DeepMind, rejected the suggestion that Google may not be trying to win the model race at all — that user reach depends more on the price-performance ratio than on standing at the frontier. Nothing is more important to the company than being at the frontier, he said, and he is fully confident Google will stay there. He then supplied no evidence for it. On Gemini 4 he repeated what was already on the record: the most ambitious launch Google has attempted, and the work is progressing if all goes well. On Gemini 3.5 Pro he added nothing, and the model appears to still be in development, already several months past its expected timeline.
Koray Kavukcuoglu on frontier models, coding agents and building AGI
Source: the-decoder.com
What he did talk about was Flash. Kavukcuoglu described the progression from Gemini 3.5 through 3.6 to 3.7 as a move from a language model to a coding agent — the intent being to turn the system from a model into an agent outright. Software engineering, he said, is the key domain and the environment in which these systems have to succeed.
Set those two halves of the conversation next to each other and the shape of Google's year becomes visible. The line that carries the frontier claim, Pro, is late and undiscussed. The line that has version numbers moving in public, Flash, is being rebuilt around a single job. A leader who wanted to argue frontier leadership had one obvious way to do it — a capability, a date, a score — and instead gave a roadmap for the other series.
I read this as a defensive appearance, not a confident one. The rebuttal Kavukcuoglu offered was to an argument nobody needs him to answer: whether price and performance matter more than frontier capability is a question for Google's commercial strategy, not for the head of its research organization. The thing only he can settle — where Gemini 3.5 Pro is and why it slipped — went untouched. Repeating an existing line about Gemini 4 is not news, and presenting it as reassurance in the same breath as a claim of frontier leadership invites exactly the reading it was meant to prevent.
The more interesting part is the substance he did give, because it quietly redefines what winning means. If the goal for the Flash line is an agent that succeeds at software engineering, then the measure stops being whether the model is better than a rival's on a benchmark and becomes whether the system finishes the task. That is a harder bar and, to my eye, a more honest one. Choosing to describe the 3.5-to-3.7 arc in those terms is a bigger statement than the frontier-leadership assertion that made the headline.
What the appearance never addressed is what frontier leadership is supposed to be measured by. Kavukcuoglu asserted it as a priority and as a certainty, without naming a model, a date or a capability that would demonstrate it. Meanwhile the one concrete commitment he made is that Google's fast-iterating series is being converted into a coding agent — which means the company's near-term evidence of progress will arrive from the line he says is not where the argument lives.