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

A Persian-language comparison of Opus 5.5 and Gemini 4 Argon on a 3D environment coding task finds Gemini clearly behind.

A day-2 entry in a 365-day AI news video series reports Google's Gemini 4 Argon release at 0:18, followed by brief items on Anthropic, Nvidia, OpenAI and others.

After Gemini 4 Argon's launch, Polymarket puts its odds of being the best model by end of October at 69%, and the author weighs an Anthropic response and a possible Claude Fable 5.5 leak on Azure.

A developer testing Gemini 4 Argon on high reasoning found it weak at one-shotting 3D worlds until Claude Opus 5.5 and GPT-6.1 Sol reviewed the output, and complained that Pro quotas are too low.

A post says Argon watched a person play cornhole, worked out the physics and wrote its own simulation for the robot to practice, which then landed 7 shots in a row and cut weeks of training to hours.

A post with a chart claims Gemini 4 Argon beats Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and Grok 4.7 on benchmarks, without giving specific scores or sources.

An Indonesian user says Google One still advertises Gemini 4 Argon as an AI Pro feature while Antigravity reserves it for the pricier AI Ultra tier, calls it a bait-and-switch, and threatens to cancel and switch to Claude Code.

A user tested Gemini 4 Argon's prose with a cultivation-web-novel breakup plot, casting ChatGPT, Claude and Gemini as sect characters and adding Jensen Huang and a Blackwell golden core as a cheat.

A developer iterated for about 30 hours with Gemini 4 Argon, Three.js and Blender to build an interactive 3D humanoid robot demo, starting from a robotic hand.

According to one user, Gemini 4 Argon watched a single cornhole video, wrote its own physics simulation, and Apptronik's Apollo robot then sank seven throws in a row.

The online benchmark MazeBench announced that a Gemini 4 Argon entry is coming soon, noting that it keeps triggering the biology filters.

A thread claims Argon has a one-million-token output limit, can solve large problems fully autonomously, and lists ten capabilities.