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

A user found Gemini 4 Argon's code review thorough and focused on changes that genuinely matter, unlike Codex's nitpicking, and Opus 5.5 agreed with its findings, though large-project development is untested.

A roundup of Gemini 4 Argon highlights: 52.8% on CursorBench, an Artificial Analysis intelligence index matching Astra at about 40% lower task cost, a 15% hallucination rate, and a pelican-riding-a-bike test.

A post with an image asks why Gemini 4 Argon is being hyped as today's mythos, reflecting the community's intense anticipation for the unreleased model.

A user shared Joke Bear and Mogura Croquette fan art made with mobile Gemini 4.0 Argon and praised the image quality on the phone version.

Benzinga relays Google's claim that Gemini 4 Argon lets robots learn physical skills by watching ordinary videos, with Apptronik's Apollo humanoid sinking seven straight cornhole shots after watching footage.

A user used Gemini 4 Argon to make a 60-second launch video for BookMyProfile, the SaaS they just launched, and shared the result.

A teaser says Gemini 4 Argon, Jev 1.14, GPT-6 Astra and Claude Opus 5.5 will compete against humans in an experiment.

The author argues that Muse Wonder and Gemini 4 Argon look strong on benchmarks but did not advance the frontier, and that the real leap will come from Fable 5.5.

A user notes that Grok 4.7, GPT 6 Sol, Muse Wonder and Gemini 4 Argon all peaked at launch and were called benchmaxxed after testing, while Opus 5.5 hype keeps going.

A user posts a screenshot saying Gemini 4 Argon perfectly reproduces the incoherent style of early models like GPT-2.

A user had Gemini 4 Argon render a described scene through Blender MCP; it is on its third attempt at the horse model and can iterate indefinitely, though it takes days.

A comparison chart lists OpenAI GPT-6.1 Sol, Claude Opus 5.5, Gemini 4 Argon, Grok 4.7, Meta Muse and DeepSeek V4.1 as the six current contenders.