Gemini 4 Pro Just Leaked: The Results Are Insane

Link to our newsletter: https://bitbiased.ai/ A mysterious Gemini model started producing surprisingly advanced interactive games and SVG demos inside Arena — while being labeled as the already-released Gemini 3.8 Flash. Just days earlier, an X account claimed to have leaked Gemini 4 Pro under the internal codename “Argon,” with a massive 2-million-token context window and 256,000-token output limit. So did Google quietly put Gemini 4 Pro into Arena for public testing, or is the internet connecting dots that were never connected? The mystery started with unusually capable Arena outputs: an interactive pelican riding a bicycle, a teapot with working controls, and eventually a playable Ultraman-versus-Godzilla browser game complete with animations, enemies, lasers, and a functioning score counter. Every one of these outputs appeared under the Gemini 3.8 Flash name. That immediately fueled speculation that Google was disguising an unreleased Gemini 4 checkpoint. And the timing made the theory even more compelling. Three days before the Arena sightings, the X account Lentils80 posted what it claimed was the first leaked output from Gemini 4 Pro “Argon.” The alleged model reportedly had a 2-million-token context window and 256K output ceiling — dramatically beyond the documented Gemini 3.8 Flash limits. But there’s a major problem: none of those Argon specifications have been independently verified. Google has confirmed that Gemini 4 exists only in a much narrower sense. Sundar Pichai said in July that Google DeepMind had begun its “most ambitious pre-training run yet for Gemini 4.” Google did not confirm a Gemini 4 Pro model, the Argon codename, those leaked specifications, or a release date. There’s also precedent for fake Gemini 4 leaks. An August “Gemini 4 Pro benchmark” supposedly showed extraordinary scores, including 99.9% on ARC-AGI-3. Those numbers were later found to match published results attributed to a competing OpenAI model, suggesting the supposed Gemini benchmark had simply been relabeled. Then comes potentially the biggest problem for the Argon theory. At least one Arena tester reportedly reproduced the mysterious output quality by selecting “gemini-3.8-flash-high” — a higher-thinking configuration of the existing Gemini 3.8 Flash model. There’s also a significant speed mismatch: the alleged Argon generation reportedly took 2.4 minutes, while some supposed Gemini 4 comparisons took far longer. The mysterious Arena demos could return in seconds. There’s no “Argon” fingerprint hidden in the generated code, no leaked model card, no API listing, and no Gemini 4 entry on Arena’s scored leaderboard. Still, the creative coding performance is difficult to ignore. The model produced polished, interactive experiences that surprised testers and generated comparisons with other frontier models. Something about these outputs caught the community’s attention — the unresolved question is whether that “something” is a secret Gemini 4 checkpoint, a higher-reasoning Gemini 3.8 Flash configuration, narrow optimization, or something else entirely. We break down the Arena sightings, the Argon leak, the evidence for and against the Gemini 4 theory, the previously debunked benchmark leak, and exactly what Google has — and hasn’t — officially confirmed. CHAPTERS 00:00 Is Gemini 4 Pro Already Hiding in Arena? 02:02 The Arena Sightings 04:05 The Argon Leak 06:03 Is This Actually A New Model? 08:42 The Case For And The Case Against 11:19 What We Still Don't Know #gemini4 #gemini #googledeepmind #ai #artificialintelligence