Gemini 4 is getting closer, but Google is behind in the AI race

What's new

For the first time, Google has put into words how far along its next big AI model, Gemini 4, actually is. At an AI conference in the US on September 23-24, Google's new DeepMind chief Koray Kavukcuoglu said that Gemini 4 is now in what the company calls "post-training." This means the base model itself has already finished training, and the engineers are now polishing safety and the ability to follow instructions before the model is released to regular users. Kavukcuoglu said Google will "roll out an early post-training version as quickly as possible," and that he hopes for a launch well before the end of the year. However, no concrete date has been set, and rumors of October or November are just guesses.

What's clever

That a model is already in post-training means the heavy and expensive work of building the "brain" itself is over. It is the same phase the competitors have gone through, but Google is emphasizing that it wants to roll the model out early and improve it continuously, instead of waiting for one big, perfect launch. It looks like an attempt to keep up with the fast pace that OpenAI and Anthropic have set with their latest models.

Cheaper or better?

Here we have to be honest: Google has published neither prices, benchmarks (standardized tests that measure an AI model's abilities) nor other technical details for Gemini 4. What we can say is that Google's current model, Gemini 3.8 Flash, is some way behind according to independent measurements. It scores 41 points on the Artificial Analysis Intelligence Index, while Anthropic's Claude Opus 5.5 gets 58 points and OpenAI's GPT-6 Astra gets 53 points. On a coding test called Terminal-Bench 4.0, Claude Opus 5.5 achieves 66.4 percent correct solutions versus 57.9 percent for GPT-6 Astra. So that is the gap Gemini 4 has to close, but whether it succeeds, and what it will cost, we will only know when the model is actually launched.

What it's good at

Since Gemini 4 does not yet exist as a finished product, no one can say for sure what it will be especially good at, or whether it will be free, paid or available in Denmark and the EU from the start. But the pressure is clear: Google currently lacks a top model that can compete with its rivals' best in areas such as coding and reasoning, and that is presumably exactly the gap Gemini 4 will try to close when it eventually reaches regular users and businesses.