Gemini 4 Pro in Arena? What is known about the rumors surrounding the checkpoint

A speculation, not a confirmed release. Arena users are discussing a checkpoint that appeared under the Gemini 3.8 Flash label in some comparisons. Based on its behavior, some infer that it is an early version of Gemini 4 Pro. Google has not confirmed this identification, so it is more accurate to describe it as a rumor surrounding an anonymous model.

Where the Gemini 4 version came from

As evidence, people publish results showing the generation of a PlayStation SVG controller, three-dimensional scenes, and web pages. The authors of the posts compare individual responses with those of other models and note the high level of detail. Such examples may draw attention to unusual model behavior, but they do not reveal its internal identifier, size, or development status.

Arena Battle mode does indeed show two models anonymously, with their names revealed after voting. The platform may also test preliminary versions. This explains why an unknown model might appear in a comparison, but it does not prove that it is specifically Gemini 4.

What has been publicly confirmed

Google’s official catalog for the Gemini API lists models from the Gemini 3 family, including Gemini 3.8 Flash and Gemini 3.1 Pro. There is no public Gemini 4 model or API identifier in the catalog. Therefore, users should not be directed to test “Gemini 4 Pro” as a publicly available product or present assessments of its quality as measured facts.

At the time of the community discussions, the checkpoint’s name was reported in different ways. No reliable primary source for the codename, including “Baryon B” or “Barium B,” could be found. These variants should not be presented as Google’s confirmed internal name.

How to evaluate such leaks

• Separate the observable result from the inference about the model’s identity.
• Do not extrapolate the quality of a single generation to programming, reasoning, and document-related tasks.
• Wait for an official model card, API endpoint, or reproducible comparisons using identical prompts and settings.

Arena’s own help centerexplains how anonymous comparisons work. The current public list of models can be checked in Google’s Gemini API catalog. Examples of user discussions do not confirm the model’s origin.