bit.fanBit.FanWeb3 News · Betting Reviews · Market Insights
Rumors of Gemini 4 Pro testing on Arena spread, but Google has not confirmed it

Rumors of Gemini 4 Pro testing on Arena spread, but Google has not confirmed it

N
News Editor
2026-09-20 05:20:54
Rumors that Gemini 4 Pro has quietly entered anonymous testing on the Arena platform have gained traction across X, where users and AI-focused accounts have circulated side-by-side videos comparing the model with GPT-6 and Fable 5. The discussion centers on claims that an Arena option labeled "Gemini 3.8 Flash" suddenly showed much stronger 3D and physics understanding than usual, leading some users to speculate that Google may be testing an unreleased Gemini 4 Pro checkpoint behind the scenes. The most widely shared comparison clips came from the official account of AI app-building platform YouWare, which described Gemini 4 as producing striking results under the same prompts used for rival models and linked viewers to its own platform. At the same time, the evidence cited by users remains thin. Much of the so-called internal probing appears to rely on asking the model to identify itself, a method that is not considered reliable because language models can generate incorrect self-descriptions. As of publication, neither Google nor Arena has confirmed that Gemini 4 Pro is being tested on the platform. No new model announcement has been made, and Gemini 4 does not appear on the public leaderboard. The only point Google has publicly confirmed is that Gemini 4 pretraining began on July 21, with CEO Sundar Pichai repeating that message during the company’s late-July earnings call.
GoogleGemini 4 ProArenaAI modelsYouWareSundar Pichaimodel testing

Talk of Gemini 4 Pro already being quietly tested on the Arena platform has been bouncing around social media. A bunch of side-by-side videos helped fuel it, pitting the model against GPT-6 and Fable 5 under identical prompts. But at the time of publication, neither Google nor Arena had backed up the claim. No fresh model reveal has happened, and Gemini 4 still is not listed on the Arena leaderboard.

The videos getting the most traction came from the official account of AI app-building platform YouWare. In those posts, YouWare lined up Gemini 4 against GPT-6 on the same prompt set, called Gemini 4’s results eye-catching, and added links pushing users to test the experience on its own platform.

In a Sept. 19 post, YouWare wrote: "Insane!!! Gemini 4 Pro shocked me again today. One prompt. Two mechanical butterflies. Same Arena prompt vs. Fable 5.1 Max. Gemini 4 Pro: ~10 min. Fable: ~30 min. Glowing wings. Fine clockwork. Gorgeous motion. You? Which journey wins? What will Gemini 4 Pro create…"

How the "ghost test" rumor started

The rumor seems to trace back to a Sept. 17 X post from user @NFT_Chen. That post said the "Gemini 3.8 Flash" option in Arena blind testing had suddenly gotten far smarter, with a big leap in 3D spatial and physics understanding, while needing 2 to 3 minutes to generate a car scene.

The same post also shared what it described as internal probing results: declared_model showed 3.8 Flash, runtime_engine showed Argon-B02, and the checkpoint identifier was G4P-INT-ARGON-SEPT11-B02. The article said a checkpoint is a saved snapshot from model training and is used only for internal testing.

That same day, another account, @LuminaBench, said it had spotted the same checkpoint and claimed it produced a "voxel pagoda" in about 5 minutes. Other accounts reportedly tried the model with a pelican SVG or ran a script involving 1,000 planets, which pushed the browser to its character cap and crashed it.

@NFT_Chen wrote in that Sept. 17 post: "Big news! Gemini 4.0 Pro has started ghost testing in Arena! Community testing suggests the 'Gemini 3.8 Flash' label may actually be tied to a 4 Pro checkpoint with the internal codename Argon. Key evidence: Arena Battle Mode randomly hits '3.8 Flash,' and its image quality/3D ability suddenly far exceeds normal Flash; internal probing shows declared_model=3.8 Flash, runtime_engine=Argon-B02…"

Why the evidence remains weak

Here’s the issue: most of these alleged internal probes seem to come from asking the model what it is. That kind of self-identification from a language model is shaky. Maybe worse. Those strings could just be hallucinated text, not hard evidence.

There is also extra attention on YouWare. It has posted the biggest pile of comparison videos while sending each post back to its own product pages, without saying why it would have access to the model in the first place. Google has anonymously tested unreleased models on Arena before, and the company has already said Gemini 4 is in training. Even so, the claims moving around right now cannot be verified on an independent basis.

What Google has actually confirmed

Google said on July 21 this year that it had started what it called its "most ambitious pre-training yet" for Gemini 4. CEO Sundar Pichai repeated that same point during the company’s late-July earnings call.

Past that, Google has said nothing about a release date, pricing, model lineup, or benchmark results for Gemini 4.

So until Google shares more, the identity of the unusually capable "3.8 Flash" appearing on Arena is still unknown outside the company.

This article was originally published by Bit.Fan. For more cryptocurrency news and market insights, visit www.bit.fan.
2400

Disclaimer:

The market information, project data, and third-party content displayed on this platform are for industry information sharing only and do not constitute any form of investment advice or return commitment.

Cryptocurrency trading carries high risks. Users should fully assess their risk tolerance and make independent decisions. All profits, losses, and legal responsibilities are borne by the users themselves.