Gemini 4 enters post-training, aims for early launch

What's new

Google DeepMind has confirmed that the upcoming AI model Gemini 4 has now entered what is called "post-training." This is the phase in which a fully trained model is refined so it gets better at following instructions, declining dangerous or unsafe requests and behaving consistently in practice. Google began the core training of Gemini 4 on July 21, 2026, and the fact that the model is ready for fine-tuning barely two months later is described by several observers as an unusually compressed timeline. The news came from Koray Kavukcuoglu, who was just promoted to head of Google DeepMind, at a conference on September 23-24, 2026.

What's clever

According to Kavukcuoglu, Google will release "an early post-training version as quickly as possible" instead of waiting for one big, polished launch. This means users may get access to parts of Gemini 4 as early as fall 2026 rather than at New Year. According to Google, the model will be significantly larger than the current Gemini models and is built especially for coding (writing and fixing computer programs) and for AI agents, that is, AI systems that can carry out several steps of a task on their own without constantly asking the user for permission.

Cheaper or better?

No price has yet been set for Gemini 4, and Google has not published official benchmark figures (standardized tests that measure how "smart" an AI model is) for the new model. On the other hand, measurements from the Artificial Analysis Intelligence Index show that Google's current top model, Gemini 3.6 Flash, scores 34 points, well behind competitors' newest models: Anthropic's Claude Opus 5.5 is at 57.6 points and OpenAI's GPT-6 Astra at 52.7. That pressure is part of the reason Google is now rushing Gemini 4 forward. Right now, Google's paid AI subscriptions cost between roughly $20 and $200 per month depending on how much you want to use the models, but it is unclear whether and when Gemini 4 will be included in those plans, and at what price.

What it's good at

Although precise figures are missing, Google itself points out that Gemini 4 is meant to be especially strong at coding and at AI agents that can work independently over longer periods, for example planning and carrying out several tasks in a row without constant human intervention. That makes the model interesting for companies that want to use AI to automate IT work or other complex workflows. For the ordinary user, the news mainly means that a significantly stronger Gemini model is on the horizon before the end of the year, but for now you will have to be patient and wait for concrete prices and independent tests.