Google plays catch-up as Gemini 4 enters testing after rivals' double launch

Google is racing to get Gemini 4 out before the end of the year, after watching Anthropic and OpenAI both ship new flagship models on the very same day, 22 September, while its own next major release remained stuck in testing.

DeepMind chief Koray Kavukcuoglu said at The Information's AI Agenda Live summit on 23 September that Gemini 4 has entered post-training and that Google wants it out well ahead of the end of 2026, though he set no firm date, per 9to5Google.

The model is already running Google's internal coding tool, Antigravity, giving Google real-world testing ahead of any public release, and a working product to point to even before Gemini 4 formally ships to customers.

The comments carry extra weight because Google's last attempt at a flagship jump stalled: a Gemini 3.5 Pro model, announced at I/O in May, never actually shipped, leaving Gemini 3.1 Pro, released in February, as the company's most recent full flagship update.

Kavukcuoglu said Google "took a little bit of a step back" to concentrate on smaller Flash models instead, including Gemini 3.8 Flash in September — a deliberate trade of flagship pace for smaller, faster releases while the bigger model matured.

The remarks were Kavukcuoglu's first public appearance since taking over DeepMind's leadership in August, when Demis Hassabis moved up to chair the unit and became Alphabet's chief scientist, making this Google's clearest signal yet of where its new leadership wants to take the model line.

Kavukcuoglu expressed confidence that Google will keep pace at AI's cutting edge when asked whether it had fallen behind, The Next Web reported. In analysis terms, the remark was aimed as much at customers weighing Anthropic's and OpenAI's recent momentum as at the audience in the room.

No model ID, pricing or model card has been published yet, leaving analysts to judge the claim largely on timing and Kavukcuoglu's own word, rather than on any concrete product detail Google has released so far.

The leadership change adds another layer to the story: Kavukcuoglu is effectively making his first public promise as chief executive, and a delayed or underwhelming Gemini 4 would land as a mark against the new regime rather than the Hassabis era it succeeded.

Investors and enterprise buyers evaluating which AI provider to commit budget to now have three recent data points to weigh: two shipped flagship models from Anthropic and OpenAI, against one still-unreleased model from Google backed only by an executive's stated intention.

That asymmetry is precisely what Kavukcuoglu's remarks were aimed at closing, even without a product to show for it yet, by putting a public timeline and a concrete internal use case behind a model that had otherwise gone quiet since May's aborted Gemini 3.5 Pro announcement.

Whether that reassurance holds will depend entirely on execution: a second stalled flagship, after Gemini 3.5 Pro, would be a far harder story for Google to spin than a first one.