When a frontier lab talks about a model that is not finished, it usually does so by accident. Google just did it on purpose. Speaking on 23 September, DeepMind senior vice president Koray Kavukcuoglu said Gemini 4 has entered early post-training and that the company intends to "roll out an early post-training version as soon as possible, because we've already seen promising results." No benchmark scores came with that. No parameter count, no context window, no price sheet, no launch date. The entire announcement was a stage in the pipeline and a promise to hurry.
To see why that is odd, it helps to know what post-training is. Pretraining is the long, expensive phase where a model learns raw patterns from enormous datasets. Post-training is everything after: reinforcement learning from human feedback, safety tuning, red-teaming, and the many smaller passes that turn a raw model into something a company will put its name on. It is not one event but a sequence of checkpoints, each more polished than the last. What Kavukcuoglu described is Google shipping from an early checkpoint rather than waiting for the fully tuned, fully adversarially tested version most labs save for launch day.
This is a genuine break from Google's own habits. Gemini 1 through 3.8 each arrived as finished products, with names, documentation, and published numbers on day one. Google was, if anything, the most cautious of the big labs about launch readiness. It is also worth remembering how the company got here: the Bard debut years ago was mocked for factual errors, and the Gemini line rebuilt that credibility release by careful release. Gemini 4 is the first entry in that arc where Google is openly trading some of the polish for speed.
The pressure behind the choice is not hard to read. OpenAI has shipped GPT-6 variants at aggressive prices, Anthropic cut costs with Claude Opus 5.5, xAI keeps iterating on Grok, and DeepSeek retires open-weight models within the same month it launches them. A multi-month gap that barely registered in 2024 can now move enterprise contracts and developer mindshare. Kavukcuoglu's comments read as a bet that the reputational risk of shipping something rough is smaller than the business risk of shipping late. The pretraining run reportedly began only on 21 July, which makes "well before the end of 2026" an unusually fast turn.
The part no one has answered is safety. An early checkpoint has, by definition, spent less time being hammered with the adversarial prompts that surface jailbreaks and confabulation before the public does. That matters more for Google than for a small lab, because Search, Cloud, and enterprise deployments all sit on top of Gemini. If an early version ships, millions of users effectively become the last mile of red-teaming. For developers, the practical takeaway is unglamorous but real: pin your model versions, keep fallback logic, and expect behaviour to shift between releases. Google is promising to iterate in public. That is another way of saying the finishing work will happen where everyone can see it.