Gemini 4 Pro leaked? Google's secret model may already be stronger than Astra and Fable
Reports from China indicate an intriguing possibility:
An anonymous model named "Gemini 3.8 Flash", which appeared on the Arena platform, may actually be an early version of Google DeepMind's next flagship model - Gemini 4 Pro.
And the numbers? If they are real, this is a serious leap forward.
The model allegedly received: 88% in DeepSWE v1.1 - coding capabilities
2,064 Elo in GDPval-AA v2
95.3% in Terminal-bench 2.1 - working with terminals and development tools
86.8% in OSWorld-2.0 - performing complex tasks on a computer.
These results are significantly higher than the results Google published for Gemini 3.8 Flash, which were 73.7%, 1,545, 89.4%, and 59.0% respectively.
And that's exactly what set the internet ablaze.
In user tests, the model allegedly demonstrated particularly strong capabilities in creating user interfaces, SVG graphics, 3D, and game development.
That is, not just "writing a better answer", but performing real work using tools - precisely where the new generation of AI models is trying to break through.
What's even more interesting are reports that Google is already working on a recursive self-improvement loop, RSI, and has simultaneously advanced the early learning phase of Gemini 4.
If this is true, the implications could be far beyond just another model upgrade.
The race is shifting from models that understand text to systems capable of planning, using tools, performing tasks, and improving themselves.
If the data discrepancies prove to be correct, Google could enter the next round of the model war with very heavy weaponry.
And most importantly - the AI race has not stopped at all - it has simply moved to a stage where the question is no longer who knows more, but who knows how to do more.
And we are just waiting to hear how many GPUs trained the new model...