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

Claude and GPT hit the prompt cache every turn (228/228) while Gemini 4 Argon hit only 167/228, and the 27% miss rate made agent loops cost up to 1.48x a perfect cache.

Users report Gemini 4 Argon is available in AI Studio to all Google AI Pro subscribers including trials, yet the same accounts still lack access in Antigravity.

With the same prompt, Gemini Argon via Antigravity CLI finished about six minutes faster and cost roughly $2.25 versus $6.63 for Claude Opus 5.5, though the author still prefers Opus's visuals.

Bundled Pro accounts (Pixel, Jio, Student) are tagged as trial/promotional and get HTTP 403 #3501, blocking Gemini 4 Argon and Claude 5.5; after Nov 2 removals only Gemini 3.8 Flash remains.

A user shows a 30-minute task built with Gemini 4 Argon and calls the result solid.

The post claims Gemini 4 can redesign a garden like a $150/hour landscape designer from one photo of a garden or balcony using seven prompts.

A short teaser: the author is building something with Gemini 4 Argon in Antigravity and says it looks good so far.

Given the same prompt, Grok 4.7 rendered a plush Elon Musk while Gemini 4 Argon produced a wooden figure, leading the author to prefer Grok.

Both models ran the Pelican-on-a-unicycle SVG benchmark at High effort, with the author posting the side-by-side results for comparison.

A video puts Claude Opus 5.5 head-to-head against Gemini 4 Argon.

A developer used Gemini 4 Argon to create a Subway Surfers-style endless runner starring a flying Tesla Roadster, with a playable link.

A model-selection list puts Claude Opus 5.5 on coding, GPT-6 Astra on review, and expects Gemini 4 Argon to handle Japanese writing and rewriting while questioning its differentiation.