The artificial intelligence industry is currently abuzz with rumors after a highly unusual event on a public testing platform. Over the weekend, developers noticed strange behavior from a model on the LMSYS Chatbot Arena, sparking speculation that Google has secretly deployed an early build of a massive, unannounced frontier model—potentially the long-rumored Gemini 4.

The “Flash” That Wasn’t

The mystery began with a model labeled “Gemini 3.8 Flash.” On the surface, this is unremarkable; Google publicly released the actual Gemini 3.8 Flash just weeks ago, on September 2nd. The “Flash” tier is specifically designed for speed, engineered to deliver fast, approximate answers rather than deep, complex reasoning.

However, users interacting with this specific Arena model noticed it was behaving entirely out of character.

When prompted to generate highly complex raw vector code (SVG) for 3D objects, the model didn’t stream a fast, vague response. Instead, it went completely silent for up to 10 minutes. When it finally responded, it produced thousands of lines of precise, perfect code, rendering objects like a PlayStation 5 and an Airbus H145 helicopter with flawless lighting, depth, shading, and geometric alignment.

This extended processing time is known as “test-time compute,” indicating a model that spends minutes internally planning, executing, debugging, and polishing before showing the user a single word. This design philosophy is completely contradictory to the “Flash” speed tier, leading developers to conclude that Google is using the “Gemini 3.8 Flash” name as camouflage for a much more powerful, slower, reasoning-heavy model.

Benchmark Leaks and Unprecedented Specs

As developers probed the anonymous model, alleged benchmark scores and backend specifications began to leak across the tech community. If true, the numbers represent a significant leap over current top-tier models like OpenAI’s Astra and Anthropic’s Fable 5.1.

• Coding and Agentic Operations: Unverified benchmarks claim the model hit 88% on deepsw SWEV1.1 (measuring agentic coding), clearing Astra by two points. It allegedly scored 95.3% on Terminal Bench 2.1 and 86.8% on OSWorld 2.0 (measuring computer use), beating both Astra and Fable 5.1 in a clean sweep.
• Massive Output Window: Researchers tracing the backend claim to have found a 10 million token input ceiling and an astonishing 256,000 token output ceiling. For context, most current frontier models cap their generation between 4,000 and 16,000 tokens. A 256k output budget would theoretically allow the model to rewrite entire software repositories or output long-form novels in a single, continuous generation.
• Physical Integration: One of the most surprising alleged backend features is native support for robotic motor control protocols—not just a text API bridge, but actual motor control translation. This strongly points to DeepMind’s ambition to build unified models intended to run directly on physical robotics hardware.

Pricing Disruption

The leaked pricing details are equally aggressive. Rumors suggest this new model will be priced at $2.25 per million input tokens and $11.25 per million output tokens.

To put this in perspective, the competing flagship models—Astra (launched Sept 3rd) and Fable 5.1 (launched Sept 1st)—both sit at $10 for input and $50 for output. If the leak holds true, Google is poised to undercut the entire frontier tier, charging roughly 22.5% of what its main competitors demand.

The RSI Rumor: Did the Model Train Itself?

The backdrop to this leak is Google’s challenging year at the top tier of AI. The company has not released a true “Pro” series update in seven months. After reportedly scrapping the Gemini 3.5 Pro architecture entirely because it couldn’t outpace the Flash tier, Alphabet CEO Sundar Pichai stated on a Q2 earnings call that the company’s focus was squarely on Gemini 4 to remain competitive.

The current rumor mill suggests that Gemini 4 finished its pre-training ahead of schedule due to a breakthrough in Recursive Self-Improvement (RSI). RSI refers to an AI agent continuously improving its own strategies and carrying that experience forward. While realistic, continuous RSI was previously estimated by industry researchers to be achievable around 2027 or 2028, rumors claim DeepMind may have successfully closed an RSI loop during the training of this new model.

While Google has not officially confirmed a model card, parameter count, or any details about Gemini 4, the tech industry is watching the Arena leaderboard closely. Consensus guesses point to a potential official reveal at a developer event later this fall, most likely in October.