Gemini 4 on the way: Google aims for launch in 2026

What's new

Google has officially confirmed the name Gemini 4 for the first time. According to the coverage, the so-called pre-training (the first, big training round) was started in July 2026, and Google DeepMind itself calls it its most ambitious training run to date. Now the model is in post-training. That is the phase where engineers fine-tune how the model behaves and test its safety.

DeepMind's chief, Koray Kavukcuoglu, has said that Google wants to get an early version out as quickly as possible. The goal is to share it before the end of 2026. However, that is a hope and not a fixed date. At least one outlet writes that a version appeared under the name Gemini 4 Argon on September 30. I have not been able to confirm that with Google, so take it with caution.

The news comes at the same time as two big stories from competitors. Anthropic, the company behind Claude, has filed confidentially to go public in the US after being valued at 965 billion dollars following a new investment of 65 billion dollars. OpenAI has, according to reports, also filed for an IPO and is aiming for up to one trillion dollars. I could not find support for the figure of 1.4 trillion dollars in the headline.

What's clever

Sundar Pichai, Google's CEO, describes Gemini 4 as a significantly larger model that is meant to be really good at writing code and working as an agent. An agent is an AI that can solve a series of tasks for you on its own, for example finding information, filling out forms or fixing bugs in a program, without you having to guide it step by step.

Google also wants to release early and improve continuously instead of waiting for everything to be perfect. For users, that could mean faster access and more frequent updates.

Cheaper or better?

That cannot be determined yet. There is no released model, no documentation and no API access, so prices per month or per million tokens (pieces of text) have not been disclosed. There are also no official measurements of speed, context length or benchmarks, that is, standardized tests.

From Google's side, the pressure is clear. The media write that Google lags behind OpenAI and Anthropic in several comparisons and that Gemini 4 is meant to close the gap. Whether it succeeds, we will only know once the model can be tested independently.

What it's good for

Google itself points to two areas: programming and AI agents. These are also the areas where competitors are currently fighting hardest, because companies pay well for AI that can relieve developers and take over routine tasks.

For ordinary Danes and companies, the advice is to wait. It is unclear when Gemini 4 will come to Denmark and the EU and whether it will be free or cost money. As soon as Google publishes prices and figures, you can assess whether it is worth switching to from Gemini, ChatGPT or Claude.