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

A user jokes that Gemini 4 Argon deserves the award for worst AI model launch and asks whether it should go to a vote.

The author says Claude Opus 5.5 took three attempts and $114 to build a honey keyboard, while Gemini 4 Argon did it for about $1, 102x cheaper.

A user says Gemini 4 Argon in Antigravity used /browser to open Chrome, download forms, merge scans into a 17-page PDF and file a full hospital insurance claim.

A blogger describes buying cheap files on Xianyu, feeding them to Gemini 4 Argon's large context to strip ads and reformat into a clean how-to manual, then selling or using it for lead gen.

A user reports the Taiwan region student offer still lets you upgrade to AI Pro cheaply, extending membership into 2027 and unlocking Argon in Antigravity.

A developer has Gemini 4 Argon build an interactive Three.js WebGPU Wembley Stadium with 90,360 clickable seats, calling it frontier-level.

A user on a paid Pro plan used the Argon model in Gemini to generate an app, then embedded it in a Google Sites page, sharing the result.

Paying Google AI Pro users report Gemini 4 Argon runs out after a handful of prompts, while the help page only says Pro gets 4x standard limits without disclosing the baseline.

A designer used Argon in Antigravity with Nano Banana to generate movie figurines and build a site tracing 100+ horror and sci-fi works across 7 eras.

The author argues Gemini Argon explains its own steps in plain language, making agent mechanics easier for beginners to learn than terser models.

The author says OpenAI's Astra beats Gemini 4 Argon on a next-action prediction benchmark, while Argon crushes Astra on long-horizon understanding and labeling.

Gemini 4 Argon briefly appeared in Antigravity for AI Pro student accounts before being pulled; it is currently only available directly through the Gemini app.