Google DeepMind is preparing an early version of Gemini 4 for release “as soon as possible,” according to Koray Kavukcuoglu, the division’s new leader. The model is in its refinement stage, he said in an interview with *The Information*, with a launch targeted well before year-end.
That is a meaningful timetable signal from Google, but not yet a product specification. The company has not detailed Gemini 4’s capabilities, pricing, deployment options, or benchmark results. Operators should treat it as an imminent platform transition to plan for, rather than a reason to change production model choices today.
What changed
Kavukcuoglu’s comments are his first reported media appearance since taking over DeepMind following Demis Hassabis’s August departure. He described Google’s intended release as an “early post-training output,” followed by fast-paced iterations.
The wording matters. Post-training is the work that typically shapes a foundation model into a useful product through instruction following, safety measures, tool use and other behavioral tuning. An early output plus rapid iteration suggests Google may be prioritizing getting a new flagship into users’ hands and improving it in public, rather than holding for a single large release.
Google has not shipped a new flagship Gemini model since the Gemini 3 series in November 2025. A Gemini 3.5 Pro update that CEO Sundar Pichai had said was coming in June did not arrive. Kavukcuoglu said the company instead stepped back to focus on faster, less capable Flash models.
Why this matters to businesses
Google is trying to resolve two different product needs at once: frontier capability and economical throughput. Flash models matter for high-volume tasks—summarization, classification, extraction and conversational features—where latency and unit cost can determine whether an AI workflow works commercially. But advanced reasoning, coding and agentic workflows often drive demand for a flagship tier.
A Gemini 4 release could therefore affect teams using Google Cloud and Vertex AI, as well as organizations building around Gemini in Workspace or Google’s developer ecosystem. The practical questions will be less about a model-version label than whether the new model improves reliability on company-specific tasks, supports required data controls, and produces acceptable costs and response times.
It also puts pressure on procurement teams to avoid overly rigid model commitments. OpenAI and Anthropic have released newer flagship lines since Gemini 3, according to the report, and Google’s next cycle may narrow or alter those comparisons. Teams that can swap models behind an abstraction layer, maintain representative evaluation sets, and track cost per successful task will be better positioned than those tied to a single vendor workflow.
What to watch next
The first release should be judged on operational evidence, not Google’s assertion that it remains at the frontier. Watch for:
- **Availability:** whether Gemini 4 arrives broadly through Google’s enterprise and developer channels, or initially in limited products and regions.
- **Product tiers:** how Google separates flagship capability from Flash-oriented speed and cost, and whether the models can be mixed in a single workflow.
- **Tool and agent performance:** concrete improvements in coding, long-running tasks, tool calling and grounded retrieval are more consequential for builders than general benchmark gains.
- **Governance details:** enterprise controls, data handling terms, auditability and evaluation tooling will shape adoption as much as raw model quality.
- **Release discipline:** the promised fast iteration cadence will be a test of the leadership transition as well as the technology.
For now, the announcement is a roadmap update, not a deployment event. The sensible response is to refresh model evaluations and integration flexibility so Gemini 4 can be tested quickly when Google provides a usable release.
