Since the release of Gemini 3 Pro late last year, Google has seen its competitors make significant progress. While OpenAI and Anthropic have released the most impressive LLMs this year, Google’s focus has been on efficient, compact models. That is expected to come to an end well before the end of 2026.
DeepMind head Koray Kavukcuoglu shared this during The Information’s AI Agenda Live Summit. Kavukcuoglu took over from DeepMind veteran Demis Hassabis. According to Kavukcuoglu, the release will come “much sooner” than the end of this year. Some Google followers are even guessing October, though that’s speculation.
According to Google DeepMind, Gemini 4 is in the post-training phase. This means that the “training run,” the resource-intensive process in which a foundation model is created based on enormous amounts of data, is complete. Post-training involves tasks such as fine-tuning and reinforcement learning, often with human feedback. Gemini 4 is therefore not yet ready; it’s also unclear whether a Pro model is on the horizon or if Flash models will be released first.
Behind OpenAI and Anthropic (and SpaceXAI, Meta, and Chinese labs)
Google was the most promising AI player between 2023 and 2025. Although OpenAI’s GPT models and Anthropic’s Claude were also able to claim the highest scores in benchmarks for a long time, Google’s overall package seemed unbeatable. It was (and is) the only player with the infrastructure, data, applications, and models to achieve full vertical integration, a luxury that even hyperscalers like Microsoft and Amazon do not have.
That this is no guarantee of success became particularly clear this year. A mature coding assistant failed to materialize, with confusing solutions like Antigravity and Gemini CLI that never quite proved to be the equivalent of OpenAI Codex or Claude Code. Meanwhile, a new dimension has been added by the fact that, since Fable and Astra, a new class of models has emerged on top of the range of LLMs with names like Opus, Sol, and Gemini Pro.
On top of that, Meta and SpaceXAI have proven that Google can be knocked off the third spot when it comes to U.S. AI performance. Grok and Muse Spark are generally more powerful than the best Gemini Flash options, while Chinese players like DeepSeek, Z.ai, and Moonshot AI have also surpassed that level.
Read also: Google Gemini 3.8 Flash can code better and more independently