Alphabet is speeding up development of its next-generation flagship AI model, Gemini 4, with plans to release the first version earlier than initially expected, potentially before the end of this year. Koray Kavukcuoglu, head of Google DeepMind, confirmed the model has entered its post-training phase, with the company aiming to ship an early post-training build soon and then iterate rapidly following release.

Gemini 4 represents Alphabet's first new flagship model since the Gemini 3 series debuted last year. While the company launched Gemini 3.1 in February and introduced several cheaper, faster Flash variants since then, its current offerings have fallen behind the latest flagship models from Anthropic and OpenAI in complex reasoning capabilities.

Gemini 4 enters post-training stage

Kavukcuoglu stated that Gemini 4 is in the early phase of post-training, a process where a foundation model, after completing its main training run, undergoes further refinement to reach a releasable standard in reliability, behavioral consistency, and real-world task performance. Google has already reviewed internal test results and plans to release an early post-training version sooner, rather than waiting for all optimization work to finish before a single full launch. Following that, the team will continue to iterate on the model quickly.

This suggests Gemini 4's release strategy may lean toward a "ship a usable version first, then upgrade continuously" approach, shrinking the gap between model training completion and market availability.

Gemini 3.5 Pro quietly shelved

Google had originally planned to release Gemini 3.5 Pro in June, a commitment announced by CEO Sundar Pichai at the annual I/O conference in May. However, that model never materialized. Kavukcuoglu explained that after shipping Gemini 3 and Gemini 3.1, Google chose to "step back slightly" and redirect more effort into the Flash model series to boost team learning and iteration speed. He did not explicitly confirm whether Gemini 3.5 Pro has been formally canceled, but he made clear that the company's focus has now shifted fully to Gemini 4.

In effect, Google has skipped the planned Gemini 3.5 Pro launch, concentrating its primary research resources directly on the next-generation flagship.

Google renews its push to close the capability gap

Gemini 4 arrives amid renewed intensification in the frontier model competition. Anthropic introduced its new Mythos model series this spring and upgraded it again this month, while OpenAI began rolling out the GPT-6 series in early September. The latest models from both companies currently surpass Google's strongest existing offerings in overall capability.

When asked about falling behind, Kavukcuoglu expressed strong confidence in the Google DeepMind team, asserting that Google will remain in the first tier of frontier model competition. Google has faced similar catch-up situations before; the launch of Gemini 3 delivered a substantial capability leap that once forced rivals to accelerate their own release cycles. As such, Gemini 4 will be a crucial test of whether Google can again narrow the frontier model gap.

Gemini 4 already used internally for AI coding tools

Google has also begun real-world internal use of Gemini 4. Kavukcuoglu said the company is developing safety guardrails and running security tests for Gemini 4, while engineers have already used the model to power Google's Antigravity AI coding tool. This indicates Gemini 4 has moved beyond lab benchmark testing into actual application stages.

By deploying the model in authentic engineering environments, Google can surface issues in complex code tasks, tool calling, and multi-step task execution early, allowing for further optimization ahead of a formal release.

Google prioritizes trustworthy agents over the AGI label

On the industry-wide discussion around Artificial General Intelligence (AGI), Kavukcuoglu said he does not focus much on whether the sector has reached that label. Instead, he emphasized that the more important question is whether the company can build agents that are truly trustworthy. This reflects a shift in some of Google's model development priorities, from simply chasing benchmark scores toward ensuring models can reliably execute complex tasks, use tools, and operate dependably in enterprise and consumer scenarios. Strengthening agent capabilities will likely be one of the key directions for Gemini 4.

Frontier model development also serves Google's TPU strategy

Gemini 4's significance for Google extends beyond software. Google is expanding its in-house AI chip (TPU) business and, starting this year, has begun selling TPUs directly to customers rather than only offering access through Google Cloud. This heightens its competition with Nvidia in the AI chip market.

Kavukcuoglu noted that as Google DeepMind develops next-generation frontier models, it can give the hardware team early insight into what computational power future models will require. This allows Google to plan the design of the next two to three TPU generations in advance, aligned with model research direction.

In other words, Gemini models and TPU chips are forming a tighter synergy: the model team provides directional guidance on future compute needs, the hardware team optimizes chips accordingly, and more efficient TPUs in turn enable next-generation model training. This represents a key structural advantage for Google over many pure-play model companies, as it simultaneously controls models, cloud computing, and its own AI silicon.

Overall, Gemini 4 is not just Google's next flagship model, but a pivotal product in its bid to reclaim the lead in frontier AI. With Anthropic and OpenAI already ahead with their latest models, Google is compressing its development and release cycles while more closely aligning Gemini 4 iterations with the TPU hardware roadmap. If Gemini 4 can arrive earlier than expected, predating the end of the year, it will serve as key evidence of whether Google can again close the frontier gap while reinforcing its integrated software-hardware advantage.