Gemini 4 is almost ready, says new Google DeepMind chief

Google is closing in on the release of Gemini 4, its next major AI model, after falling behind rival labs on flagship releases. Koray Kavukcuoglu, in his first public appearance as the new head of Google DeepMind, told The Information that the model is currently in a refinement stage — meaning it exists, it has been trained, and engineers are now polishing it before any broader rollout.

This piece covers what is known about Gemini 4's development status, what refinement means in practice for a frontier model, and what developers and enterprise teams building on Google's AI stack should be paying attention to right now.

Gemini 4 is almost ready: what actually changed

Kavukcuoglu's statement marks the first time Google leadership has publicly acknowledged Gemini 4 by name and confirmed it is in active preparation for launch. Previous communications from Google had centered on incremental updates to the existing Gemini family — Flash variants, experimental versions in AI Studio — without signaling that the next-generation flagship was imminent. The admission of a refinement stage is meaningful precisely because it places Gemini 4 past the training phase, which is typically the most compute-intensive and time-consuming part of building a large model.

Google's position relative to competitors has drawn attention inside and outside the company. OpenAI shipped GPT-4o and successive updates, and Anthropic moved through the Claude 3 series at pace. Kavukcuoglu's remarks, framed as a leadership introduction rather than a product announcement, appear calibrated to signal momentum without committing to a specific timeline.

Gemini 4 is almost ready: how it works

Refinement, in the context of frontier model development, generally encompasses a set of post-training steps applied after the base model has completed its initial training run. These steps typically include reinforcement learning from human feedback, red-teaming for safety and alignment, capability evaluations across benchmarks, and iterative tuning to improve instruction-following and reduce failure modes. None of these are quick processes — they can extend for weeks or months depending on the scale of the model and the number of issues surfaced during evaluation.

Google DeepMind operates a large research and engineering organization that runs these processes in parallel across multiple model variants. Kavukcuoglu has not specified whether Gemini 4 refers to a single model or a family of models at different sizes, which is the pattern Google used for Gemini 1.5 and its successors. Builders should expect that, as with previous generations, a flagship release would eventually be followed by lighter, faster, and more cost-efficient variants.

Gemini 4 is almost ready: why it matters now

Google's delay relative to competitors on flagship model releases has had practical consequences. Enterprise customers evaluating AI infrastructure have had to make decisions about which provider's API to standardize on, and a gap between Google's offering and the latest models from OpenAI or Anthropic can push those decisions toward competitors. Kavukcuoglu's public confirmation of Gemini 4's status is partly a retention signal — a message to developers and enterprise buyers that a significant upgrade is on the near horizon.

For Google specifically, a flagship model release also functions as the foundation for updates across its product surface: Workspace features, Search Generative Experience, and the Gemini API that third-party developers use. A materially stronger Gemini 4 would allow Google to refresh capability claims across all of those products simultaneously, giving the company a moment to reset its competitive position in a market that has moved quickly.

Gemini 4 is almost ready: who is affected

Developers building production applications on the Gemini API are the most immediately affected group. Any significant architectural or capability change between Gemini generations has historically required prompt re-engineering, re-evaluation of output quality, and in some cases updates to integration code as the API surface evolves. Teams that have built workflows around current Gemini models should begin auditing their implementations to understand where they rely on behavior that might shift with a new release.

Enterprise customers using Google Workspace AI features and organizations running Vertex AI workloads are also in scope. Google typically stages model rollouts, making new versions available in preview before general availability, which gives larger teams some runway to test. Independent researchers and benchmark-watchers will be among the first to get signal on Gemini 4's actual performance once Google begins any form of staged access.

Gemini 4 is almost ready: what to watch

The refinement stage Kavukcuoglu described does not come with a public completion date, so the clearest leading indicator will be changes to Google's AI Studio and Vertex AI preview offerings. Google has historically used these channels to surface experimental model versions before official announcements. Any appearance of a new model identifier in API documentation or developer tooling should be treated as an early signal that a formal release is approaching.

Kavukcuoglu's tenure as DeepMind leader is itself worth tracking. His choice to make Gemini 4's status the centerpiece of his first media appearance suggests it will be a defining moment for his leadership and for Google's AI credibility heading into the next product cycle. Builders who want to be ready at launch should document their current model performance baselines now, so they have a clean comparison point when Gemini 4 becomes testable.