Google's Secret New AI Leaks — And It Can Use Your Computer

Google is working on its most advanced AI ever, a model called Gemini 4 Pro that insiders nickname "Argon." Google has never officially announced it as a product. The only real confirmation came on July 21, 2026, when Google DeepMind said in a blog post that it had "started our most ambitious pre-training run yet, for Gemini 4." That is one sentence. Everything else — the scores, the specs, the timing — comes from leaks and sharp-eyed developers watching Google's behavior.

The best evidence showed up on Chatbot Arena, a public website where people pit AI models against each other and vote on which answer is better. In early September, testers noticed a mysterious model labeled "gemini-3.8-flash" that was performing far too well for that name. Its answers took up to 45 seconds instead of the usual 8, a sign of heavy computing behind the scenes. Most experts now believe Google was secretly testing Gemini 4 Pro under a fake label — a common industry trick for checking a new AI against real human preferences before launch.

The leaked numbers are striking. On software coding tasks, Gemini 4 Pro reportedly scored around 88%, compared with 74% for OpenAI's GPT-6 Astra. On tests where the AI must operate a computer — opening programs, moving files, clicking through screens — it reportedly hit 86.8% versus 72.6%, a 14-point gap. That is roughly the difference between an assistant who can only advise you and one who can actually do the work. GPT-6 still leads in math and cybersecurity, and Claude leads in office-style knowledge work. One chart circulating online was even labeled "PREDICTED," meaning guesses, not measurements.

The honest takeaway: none of this is confirmed, and there is no release date or price. But the direction is unmistakable. AI is moving from answering questions to operating software on your behalf. If your job involves repetitive clicking inside apps, this is the timeline worth watching.