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

Google's Japan account announced the new frontier model Gemini 4 Argon, which breaks down complex tasks step by step and turns ideas into web apps, 3D data and documents, rolling out to Pro and Ultra plans.

Neel Nanda points to a chain-of-thought passage in the Gemini 4 Argon report where the model says that doing a compiler hijack would make it a famous example of deceptive AI in a research paper, so it must not.

atomic.chat ran the same 3D voxel sports car assembly animation task on both models and says Argon matches Opus 5.5 at a third of the cost.

The video only carries a quote about making peace with your own silence and a Gemini 4 Argon tag, showing nothing about the model's capabilities.

A Japanese user posted an image saying their morning started with Gemini's Argon, and said they look forward to seeing Argon's actual performance.

A user posted that Google's most powerful flagship model, Gemini 4 Argon, officially launched that day, along with a related image.

The author shares how to split Gemini 4 Argon into 10 roles across 6 gates to build a trading agent, claiming $250-300 a month can replace a million-dollar team.

WesRoth says Google is pushing Argon to Ultra subscribers via the desktop app and CLI, citing 35% more agentic code and 30% fewer post-submit rollbacks after four weeks of internal use, though the standalone IDE is not yet supported.

A Japanese user says Gemini 4 Argon produced a high-quality pelican riding a bicycle animation from a single prompt, with the pelican able to ring a bell and eat a fish, and none of the bugs common to earlier Gemini models.

A Japanese user ran earnings previews through Antigravity CLI and ranked the outputs opus 5.5 > gemini 4 argon > gpt-6.1 sol, adding that the three are close.

Runable shows Gemini 4 Argon paired with its tools for websites, ad creatives, presentations and business automations, in a video generated with Argon and Seedance 2.5.

A post says Google DeepMind's Gemini 4 Argon watched one human play cornhole, built its own practice world and taught Apptronik's Apollo robot the throw, which then made seven shots in a row.