Gemini 4 Release Timeline: DeepMind Says It Is in Post-Training and Coming Much Earlier Than Year-End

September 25, 2026. Google has been the quiet one in this month's frontier race, and on Wednesday its new DeepMind chief said why, and for how long. Speaking at The Information's AI Agenda Live summit on September 23, Koray Kavukcuoglu said Gemini 4 has entered the early stages of post-training and that Google intends to release an early post-training output as soon as possible, much earlier than the end of 2026, then keep iterating, as reported by Dataconomy and Android Headlines. For the businesses we serve, the model itself matters less than the thing it powers: Gemini is the engine behind Google's AI Overviews and AI Mode, so the next version decides which companies get named when a customer asks Google a question.

What was said, and what it implies

1. Stage: post-training, the phase in which a pre-trained base model is refined with human feedback, reinforcement learning and specialised data. Pre-training is finished.
2. Timing: much earlier than the end of 2026, released in stages, starting with an early version and iterating on feedback rather than waiting for a polished launch. No date was given.
3. The gap it closes: Google has not shipped a flagship since the Gemini 3 series in November 2025. OpenAI released GPT-6 Astra on September 3, Anthropic released Claude Fable 5.1 on September 1 and Claude Opus 5.5 on September 22.
4. The miss it follows: Sundar Pichai said at I/O in May that a Gemini 3.5 Pro update would arrive in June; it never shipped. Kavukcuoglu said Google took a step back to focus on faster, less powerful Flash models.
5. The context: these were Kavukcuoglu's first public remarks since taking over DeepMind after co-founder Demis Hassabis left in August. Alphabet shares fell nearly 5% over two sessions through Wednesday as Meta's Muse agent climbed the app charts. Asked whether Google was falling behind, he said it is a certainty that Google will always be at the frontier.

Why a staged release matters more than the date

A staged Gemini 4 means Google's search features will be running on a model that changes every few weeks rather than once a year. AI Overviews and AI Mode already re-rank which sources they cite as the underlying model moves; a rolling upgrade makes that churn continuous. On our own property, Google's generative AI report shows 6,330 impressions in AI features over the last 90 days, and the pages that earn them are the same ones that earn clicks. A model change that re-weights how Gemini judges authority, freshness or entity clarity shows up in that report within days.

What it means for operators

Do not wait for Gemini 4 to build anything. Every workflow that runs on Gemini's Flash models today will run on the next generation with a model string change, and the price cuts of the last two weeks from OpenAI and Anthropic mean waiting costs more than migrating later. What is worth doing now is the work that determines whether Google's AI names your business at all: a clear entity page, consistent name, address and service data across the web, pages that answer the specific questions customers ask, and off-site mentions that AI systems treat as evidence. That is the substance of AI search optimization, and the reason to start before the model ships is that the citations Gemini 4 learns to trust will be the ones that already exist when it is trained. Watch two signals in the meantime: the Search Console generative AI report, which now also carries the multimodal filter we covered this week, and Google's own model release notes, which will say when a new Gemini version begins serving AI Overviews.