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

Pichai said on the Big Technology Podcast that Gemini 4 Argon consumed 25 trillion tokens in its first 72 hours, about 3x the prior generation, and admitted compute is short every day; the post links a 3m40s Opus-made summary.

Opus 5.5, Gemini 4 Argon, DeepSeek V4.1 and MiniMax M3.1 each built a jet dogfight video from one prompt; Argon cost $2 and took 21 minutes.

A user recounts playing 20 Questions with Gemini 4 Argon, which nailed the answer (a floppy disk) by question 10 and then gloated about it.

Google says Gemini 4 Argon helped identify and mitigate critical security vulnerabilities in PHP, highlighting AI's role in software security.

A user says Gemini 4 Carbon is already on Arena with noticeable speed and quality improvements, hoping Google ships upgrades every 1–2 months.

Sundar Pichai says Gemini 4 Argon processed nearly 25 trillion tokens in its first 72 hours, about 3x Gemini 3, while Google Cloud forgoes short-term revenue to supply DeepMind.

A short image post where the author can't decide between Gemini 4 Argon and GPT Astra, offering no hands-on results or new details.

The author guesses Gemini 4 Flash won't cost much and may be usable within a Google Pro subscription, citing pricing of Luna, Haiku and Chinese Flash models.

PromptLLM posts an image of a controversial take produced by Gemini 4 Argon, circulating it as a community meme.

A user shares a set of 3D objects created with Gemini 4 Argon, calling the results delightful and worth browsing.

A poll asks which single model to keep among Grok 4.7, GPT-6.1 Sol, GLM-5.3, Claude Opus 5.5, Gemini 4 and Argon 4.

In a 780K+ human-vote benchmark of LLMs generating 3D objects from code, Gemini 4 Argon ranks first, followed by GPT-6 Astra.