Google advances Gemini 4 to post‑training phase and begins internal trials in Antigravity

Post‑Training Shift for Gemini 4

Google announced that its Gemini 4 model has moved from the initial training stage into a post‑training phase focused on behavioral tuning.

The shift was disclosed by DeepMind chief Koray Kavukcuoglu during The Information’s AI Agenda Live Summit on September 23.

Kavukcuoglu said the company aims to release an early post‑training version “as soon as possible” and hopes to make Gemini 4 available well before the end of 2026, though no specific launch date was given.

Post‑training, according to the remarks, involves refining the model’s behavior, conducting safety tests and installing guardrails before broader availability.

The approach allows Google to continue iterating on a version that has already passed initial training while tuning and testing are still underway.

Testing Within Antigravity

Google is evaluating Gemini 4 inside Antigravity, its internal agentic software‑development platform.

Antigravity is positioned as a place for building AI agents, and using Gemini 4 there provides a practical internal use case.

The testing focuses on whether the model can power agents that developers will trust to perform work tasks.

Kavukcuoglu emphasized that “the conversation is more about are we able to build intelligent agents that we can trust,” highlighting the trust‑centric goal of the trials.

Antigravity currently serves more than 2.4 million weekly active users, according to Google’s July 22 earnings call.

By integrating Gemini 4 into a real‑world coding workflow, Google can assess the model’s behavior in a software‑development context rather than relying solely on benchmark scores.

No benchmark results or release format for Gemini 4 have been disclosed, and the internal test does not guarantee that the model will outperform competitors in coding tasks.

Implications for Developers and Google’s Roadmap

Google’s developer ecosystem already includes over 9 million monthly active developers using its model APIs, processing roughly 22 billion tokens per minute.

This existing scale underscores why the quality of Gemini 4 in an agentic coding environment is strategically important for Google.

The company’s recent model naming timeline includes Gemini 3 Pro in November 2025, Gemini 3.1 Pro in February 2026, and several Flash updates, making the flagship timeline harder to read.

Kavukcuoglu noted that Google took “a little bit of a step back” from Gemini 3.5 Pro to focus on Flash models before concentrating on Gemini 4.

Google has not confirmed whether Gemini 3.5 Pro will continue to ship alongside Gemini 4.

The internal Antigravity test represents a concrete step toward demonstrating that Gemini 4 can be trusted in agentic applications before any external developer release.

While the company has not announced external access to Gemini 4 through Antigravity, the internal trials signal a readiness to evaluate the model’s real‑world performance.

Overall, the post‑training phase and Antigravity testing illustrate Google’s intent to align Gemini 4 development with its broader developer and agent product strategy.

Why This Matters: Google’s early post‑training testing of Gemini 4 signals its intent to ship a model that can be trusted in developer‑facing agentic tools before 2026.