Google is testing Gemini 4 inside Antigravity, its agentic coding platform, before release

"The conversation is more about are we able to build intelligent agents that we can trust," Kavukcuoglu said, in a reproduction of The Information interview that The Decoder corroborated ^[3]. Runtimewire argues the Antigravity test puts that question ahead of the usual frontier-model scorekeeping: can the next flagship power agents people will trust to carry out work ^[5]. I agree with that ordering for teams that run agents. A coding agent can go wrong at any step of a long task, and a one-shot benchmark score grades one answer.

An internal run in Antigravity tells Google how Gemini 4 behaves in a software workflow Google controls ^[7]. According to Runtimewire, it does not establish that the model will beat competitors at coding, or that outside developers will get it through Antigravity first ^[7]. Google has not published benchmark results or a release format ^[8]. For Google's internal findings to predict anyone else's results, that team's repositories, tool permissions and task lengths would have to resemble the work Google's engineers give the agent.

The early release plan changes how testing has to be done. Runtimewire describes an early post-training version as a way for Google to keep iterating on a model already past initial training while tuning and testing continue ^[9]. The same report calls the plan a stated intention, not a schedule or a promise of general availability ^[9]. The checkpoint teams test first can differ from the one Google later calls finished. A pass rate on the early version measures the early version. The harness needs to log the exact model identifier with every run, so the suite can be repeated when Google swaps it.

Google's naming and release sequence has made the flagship timeline harder to read, in Runtimewire's account ^[17]. Gemini 3 Pro arrived in November 2025 and Gemini 3.1 Pro followed in February 2026 ^[11], about three months apart ^[1]. Since then Google has issued several Flash updates, most recently Gemini 3.8 Flash ^[12]. The Flash numbering has now passed a Pro version that may never ship ^[6].

Google already has developer products where a stronger model could be used ^[14]. Pichai said on the July 22 earnings call that more than 9 million developers were building monthly with Google's models and that its model APIs were processing about 22 billion tokens per minute ^[13]. Those are company-reported figures for the existing business, not Gemini 4 adoption ^[14].

Kavukcuoglu said Google remains confident it will stay at the frontier, despite questions about whether it has fallen behind OpenAI and Anthropic ^[15]. That claim will be judged once Gemini 4 can be compared and used outside Google's internal systems, according to Runtimewire ^[16].