Gemini 4 Pro / Flash: benchmarks, hands-on tests and leaks

Genie 4 adds real-time ambient audio and character dialogue, holds scenes for about two minutes, runs on Gemini 3.8 Flash, and pairs with Gemini 4 Argon for robot training.

A developer lists what one Gemini API key enables for agents—image/video understanding, audio, generation, TTS, cheap transcription and classification—and says Argon 4 is creative and uncorrelated with Claude/OpenAI models, near frontier but still behind Opus 5.5 for daily use.

Google DeepMind unveiled Genie 4, a real-time photorealistic world model steered by natural language, and pairs it with Gemini 4 Argon and Gemini Robotics to simulate Apptronik's Apollo humanoid playing checkers.

A Turkish thread details Genie 4's new capabilities—synchronized audio generation, real-time character dialogue, better human motion and physics—plus a Gemini 4 Argon-driven Apollo robot checkers simulation.

DeepMind demoed Argon breaking a task into steps and feeding text instructions to Genie 4 as it generates the scene live—a robot simulation experiment, not a physical robot demo.

DeepMind's Genie 4 world model demo has Gemini 3.8 Flash break tasks into steps and replan, plus an experiment where Gemini 4 Argon drives an Apollo humanoid playing checkers.

A team combines Gemini 4 Argon with Genie 4 for robotics: Argon splits user instructions into steps fed as live captions to steer Genie 4's simulation as a task progresses.

Google says it paired interactive world model Genie 4 with Gemini 4 Argon in DeepMind robotics research to simulate scenarios in imagined worlds.

A user comparing the same prompt and directory says Gemini 4 Argon finds more accurate sources and explains them more clearly than fable-5.1 for chatbot and consulting agent use.

A Turkish monthly AI roundup says Gemini 4 Argon launched for AI Pro and Ultra subscribers on October 13 with a background tasks mode.

Rumor says Gemini Carbon, a new Gemini 4 model built on Argon, is in early access for Google teams and partners, targeting long-horizon coding, agent workflows and 3D generation at 2–4x slower speed.

The same supercar prompt run through each lab's own agent: Gemini 4 Argon took 8m41s and $0.48, succeeding on the second try.