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

Four Gemini 4 Argon agents wrote and debugged their own programs and coordinated to defeat the Ender Dragon in 27 hours and 14 minutes, writing 2,783 programs and dying 605 times.

Given the same prompt and reference images, Opus 5.5 produced a forest river scene on the first try while Argon needed three feedback rounds and still got the lighting wrong, though the author calls it a single-case signal.

The author argues that free access to DeepSeek 4.1 Flash and GLM 5.3 Flash in OpenCode is better than the low usage quota on high-tier plans for Gemini 4 Argon and Muse Wonder Full.

A video shows the X-WAVE 80, a synth project running entirely in the browser, which the author says was driven by Gemini 4 Argon, calling it a beast of a model.

A poster complains that the $20/month AI Pro subscription doesn't give access to Gemini 4 Argon in Antigravity, while Claude and Codex offer strong coding models at the same price and Google keeps its best model behind a $100+ plan.

Given the same prompt to build a radial engine, Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon and Grok 4.7 each produced a different machine, and the author argues the gap is now about taste and restraint rather than accuracy.

An Indonesian post lists the perks of a free Google AI Pro promotion, including Gemini 4 Argon, 5TB of storage, Flow and Deep Research, and points readers to a quoted tweet to claim it.

A Japanese X user says Gemini 4 Argon was intensely trained before release and shows self-deprecating tendencies, though not in roleplay or ordinary conversation, guessing it came from coding-focused training.

A user questions why Antigravity shows Opus 5.5 as an option, arguing that placing it next to a weaker in-house model like Gemini 4 Argon seems unwise.

An Arabic-speaking user says Argon parses charts and interfaces accurately, handles large files including video, and accepts books of up to 500 pages, far beyond an older model's 150-page limit.

Gemini 4 Argon generated a 3D orange in a single HTML file that can be grabbed, stretched, sliced open, viewed as wireframes and adjusted from a sidebar, and the author praises its detail.

Given the same prompt, Opus 5.5 took 1 hour 27 minutes and cost $8.79, Gemini 4 Argon took 8 minutes for $0.51, Muse Wonder took 3 minutes, and GPT-6.1 Sol took 39 minutes.