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

A user shows Gemini 4 Argon generating raw SVG code for a pelican riding a bicycle, a quick probe of its vector-graphics output.

A developer notes that Gemini 4 Argon is accessible through the opencode Go subscription, calling frontier models cheaper than ever.

In an Arena test, Gemini 4 Carbon generated a neon restaurant scene in about 45 minutes, with results close to what Opus 5.5 produced.

A video runs the same lighthouse challenge across four models, showing Gemini 4 Argon at $6.30 versus Claude Sonnet 5.5 at $29.90, Opus 5.5 at $53.57 and GPT-6 Astra at $42.28.

A user says Gemini 4 Carbon turned a single 3D seaplane prompt into an infinite ocean with random island generation and driving controls, and calls it far stronger than Argon.

A reply asks whether Gemini 4 Flash is already at Sonnet 5.5 level, with no benchmark evidence offered.

A French post claims the new Gemini 4 checkpoint Carbon is much more powerful than Gemini 4 Argon and surpasses Opus 5.5 on 3D tasks.

MiniMax M3.1 Flash rose from 17% to 56% on AA Omniscience, passing Gemini 4 Argon; the author altered public questions and found the model still returned memorized answers, suggesting benchmark overfitting.

A user claims Google will release two models, Gemini 4 Carbon and Gemini 4 Flash, in the coming days, with an image attached.

The author previews imminent releases: Anthropic's Fable 5.5, OpenAI's GPT-6.1 Astra, and Google's Carbon plus a Gemini 4 Flash that could mark a real comeback.

Four Gemini 4 Argon agents with zero human intervention ran 2,783 programs over 27 hours and coordinated via 229 messages to kill the Ender Dragon.

A user notes Google is iterating Carbon quickly in public testing and that Gemini 4 Carbon now feels very close to release quality.