Gemini 4 is coming — here's what Google has said so far

If you use Google's Gemini assistant, or products built on it, here's the headline: Gemini 4, the company's next major model, is now in its final stages of testing, according to Google DeepMind's chief executive.

Koray Kavukcuoglu said at a conference on 23 September that Gemini 4 has entered what's called "post-training" — the stage where a company runs safety checks, sets up guardrails, and fine-tunes how the model actually behaves, before it's considered ready for the public, according to 9to5Google.

It's already being used inside Google itself, powering a coding tool called Antigravity, even though it hasn't launched publicly yet, which suggests the model is far enough along to handle real work, just not ready for a wider release.

On timing, Kavukcuoglu said Google wants to release Gemini 4 well before the end of 2026, though he didn't give an exact date, so there's no confirmed launch window to plan around just yet.

This would be Google's first fully new flagship model since Gemini 3 Pro came out in November 2025 and was followed by a smaller update, Gemini 3.1 Pro, in February. A bigger update, called Gemini 3.5 Pro, was announced back in May but never actually launched, according to The Next Web.

Kavukcuoglu explained that Google deliberately slowed its flagship releases to instead focus on smaller, faster "Flash" models, including one called Gemini 3.8 Flash that came out in September, while the bigger Gemini 4 model kept being refined behind the scenes.

The update comes soon after rivals Anthropic and OpenAI both released their own new flagship AI models, Claude Opus 5.5 and GPT-6 Sol and Luna, on the same day, 22 September, putting fresh pressure on Google to catch up in the eyes of everyday users.

For now, there's nothing new for you to try: no Gemini 4 model is publicly available, no pricing has been announced, and no official release date has been set, so the most useful thing to do is simply wait for Google's next update.

It's also worth knowing what "post-training" actually involves, since you'll likely see the term again: it covers safety checks, setting rules for what the model will and won't do, and adjusting how it responds, all separate from the much larger process of first teaching the model using huge amounts of data.

Google using Gemini 4 internally, inside its own coding tool, before letting the public near it is a fairly normal step for a company this size, and it gives Google a chance to catch problems early rather than finding them after a wide public release.

If you're deciding which AI assistant to rely on right now, this update doesn't change much yet: Gemini 4 isn't something you can use today, so whatever you're currently using, whether that's Gemini or a rival product, remains your best option until an actual launch happens.