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

Google released Gemini 4 Argon, accessible through the Gemini app, AI Studio, AI Mode in Search and Antigravity, with a focus on specialized domains.

A security researcher says they submitted their first P1 bug bounty report, found with Gemini 4 Argon in two minutes.

A user says Gemini 4 Argon is nearly as smart as Claude Opus 5.5 but cheaper and faster, and is testing it on information gain, internal link restructuring, topical authority planning and link outreach.

Argon is now available to Pro and Ultra subscribers and via API; it ties GPT-6 Astra at 53 on Artificial Analysis but trails Opus 5.5, and some employees say it underperforms its benchmarks in real development work.

Google's flagship Gemini 4 Argon is live at a promotional $2/$10 per million tokens, one-fifth of GPT-6 Astra's price, with cached input at a 5% discount and a rise to $4/$20 after November 30.

DeepMind co-founder Demis Hassabis says he is itching to vibe-code an open-world game, and the poster says one prompt on Gemini 4 Argon's launch night produced a playable 5v5 anime shooter for $0.29.

A comparison post says Anthropic has Claude Opus 5.5 and Google has Gemini 4 Argon, argues OpenAI needs a serious comeback, and concludes the AI race is far from over.

The author tested Gemini 4 Argon High on their own bench and called the results a joke, saying Gemini 3 Pro was better.

A user posted a screenshot complaining that Gemini 4 Argon's usage limit is too tight, hitting the cap after roughly ten messages.

A user asked Gemini 4 Argon to generate an SVG animation with background music; it took 70 minutes, used about 29.6 million tokens and cost $6.66.

Given the same prompt, Gemini 4 Argon finished first, GPT-6.1 Sol produced the best overall output, Muse used the most credits for the weakest result, and Opus 5.5 was slowest.

A developer names Opus 5.5 in Cursor and Gemini 4 Argon in Antigravity and Cursor as the two models they use most right now.