Google Gemini 4 Pro spotted under an alias on LM Arena

The sudden appearance of a mysterious checkpoint on the blind testing platform has sparked widespread speculation about the imminent launch of Gemini 4 Pro, setting the stage for a fierce technical showdown in the generative AI model market.

Mysterious appearance of the flagship version under an alias

On September 26, the independent AI evaluation platform based on blind comparisons, LM Arena, saw the detection of a highly sophisticated trial version that developers and independent testers believe is Google’s highly anticipated Gemini 4 Pro flagship model. Surprisingly, this version appeared disguised under the name Gemini 3.8 Flash, borrowing the identity of a lightweight and fast model that Google officially launched on September 2. This tactic is a familiar and common method used by major tech companies to test their advanced models with complete confidentiality and without revealing their true identity to gauge actual performance against real users.

This spotting comes just days after an official announcement by Google DeepMind confirming that the Gemini 4 model has entered the “post-training” phase, putting the American search giant on track to commercially launch the model well before the end of the current year, as part of its efforts to strengthen its competitive position against traditional and new rivals.

Exceptional performance and ultra-fast data processing

Specialized developer Pankaj Kumar was the first to officially draw attention to the discovered checkpoint on September 26, describing it as having an astonishing response speed that is nearly twice as fast compared to the most prominent flagship models currently available in the arena. Kumar confirmed in his initial evaluation that the model displays “the best design and organizational taste he has ever seen in any AI model,” whether in delivering complex answers or formulating programming interfaces and logical instructions.

This observation was not an isolated case or a passing event; tester reports began pouring in regarding the appearance of internal trial versions of Gemini 4 Pro since roughly September 14. A specialized tech account named Lentils revealed additional details stating that the model’s internal codename within Google’s halls is Argon. The account noted that the model possesses an exceptional capacity for input and output reaching up to 256,000 tokens at a time, representing four times the maximum limit available in the previous generation, allowing it to process and generate massive documents and code without interruption.

Repeated sightings and leaks regarding the codename Argon

Detections of additional checkpoints for the model continued on September 17 and 19 under various names borrowed from the Flash family, before the September 26 version attracted widespread attention in global tech circles thanks to its qualitative leap in output quality and logical reasoning speed. Experts believe these camouflage names aim to gather objective human evaluations free from any prior bias toward the brand or expectations associated with major core releases.

DeepMind confirmations regarding the post-training phase

In a related context, Koray Kavukcuoglu, Vice President at Google DeepMind, announced during his participation in the Live AI Agenda summit organized by The Information on September 23 that the Gemini 4 model is currently passing through its early stages of post-training and precise performance tuning. Kavukcuoglu stated clearly that the company aims to officially roll out the model at a time “much earlier” than the end of the current year.

The tech official revealed that Google engineers are already using the model internally and intensively via the company’s programming and development platform known as Anti Gravity, alongside subjecting the system to rigorous and intensive safety and reliability tests to ensure the safety of its outputs. Despite these enthusiastic statements, Google officially refrained from confirming or denying its connection to the checkpoint spotted on LM Arena, nor did it comment on the accuracy of the codename Argon. The company’s last official statement regarding this generation was issued on July 21, when DeepMind announced the start of the largest and most ambitious pre-training cycle in its history for the Gemini 4 model.

A heated October and competing against rivals

Developer Kumar suggested that if the pace of accelerated improvements shown by the checkpoints continues, Google might proceed to launch Gemini 4 Pro between early and mid-October. This expected timing puts the model in direct and fierce competition with other giant models already launched or nearing release, foremost among them Anthropic’s Claude Opus 5.5 and OpenAI’s upcoming GPT-6 Astra.

Although these timelines remain mere expectations and speculations from the developer community and have not been officially proven by Google, Kavukcuoglu emphasized that the company prefers publishing early results of the post-training phase and moving with a rapid iterative development style, rather than delaying the launch to wait for a fully polished final release, which boosts the likelihood of seeing the model very soon in users’ hands.

Frequently asked questions

Question: What fake identity did the Gemini 4 Pro model appear under on LM Arena?
Answer: It appeared disguised under the name Gemini 3.8 Flash to test its real performance and speed anonymously and neutrally.

Question: What is the model’s internal codename and what are its most prominent leaked specifications?
Answer: It bears the codename Argon and features a maximum output limit of up to 256,000 tokens and a speed twice that of the previous generation.

Question: What is the potential date for the official market rollout of the model?
Answer: Statements from DeepMind leadership and expert speculations point to a likely launch in the first half of October or before the end of the year.

Tagged , , , ,