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

A user posts an image claiming Gemini 4 Flash and Gemini 4 Carbon were spotted, hinting at upcoming Google models.

The author used Argon 4 in the Gemini App with a single prompt to generate an interactive visual demo of how a car works inside.

Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon and Grok 4.7 each build a radial engine from the same prompt, compared on cylinder count, symmetry and how much each packs in.

A user shares a ~30-minute motion design made with Gemini 4 Argon: flashy and distinctive, but with harsh colors and awkward overlapping Gemini logos.

After paying $200/month for Ultra to try Argon, the user calls it a letdown: verbose writing, weak 3D and animation versus Opus 5.5, and poor code.

A Japanese user says Gemini 4 Argon, driven through Claude Code, processed 150GB+ of vlog B-roll into a 30-second OP/ED project with BGM-synced cuts and blackout transitions.

A user details broken custom instructions, failing tools, lost task context, multi-minute delays and dropped voice input in the Gemini app, even with Argon on high effort.

In the same nuclear strike simulation, Gemini 4 Argon cost about $3 in 34 minutes while Opus 5.5 cost $148.65 over nearly 12 hours.

An unconfirmed rumor list of five models to watch includes Kimi K3.1, Qwen 4, Fable 5.5, Gemini 4 Carbon, and GPT 6.1 Astra.

NovaXCode posts an image claiming sightings of two model names, Gemini 4 Flash and Gemini 4 Carbon, hinting Google is quietly cooking new releases.

Hermes Agent now has a /print3D view that turns text or an image into a printable 3D model in minutes, with Argon recommended for modeling.

A ModelCompare experiment found 29 of 35 answers nearly identical, with Gemini 4 Argon likening human consciousness to a PR department that only spins decisions after the fact.