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

A user asks how Gemini, long treated as the laughing stock of AI, suddenly became competitive with Opus after launching Argon.

A short positive note saying Gemini 4 Argon does not treat the user like an idiot when asked something, and that its interactive interface elements look great.

The author used Gemini 4 Argon in Google AI Studio to generate marbles-music, a web page where dragging sliders and changing the scale lets you improvise music.

The author says Argon costs half of Opus 5.5 and condenses the new features, available channels, the December 1 price increase, its strengths, benchmark sources and three things to try this week into a single page.

A roundup of the entry points and costs for running Gemini 4 Argon: Google AI Pro, OpenCode Go, Devin Pro, an older Cursor subscription and Google Ultra via a Turkish subscription.

Both models drafted implementation plans for a real Python backend feature and GPT-6 Astra scored them 77/100 for Opus 5.5 xhigh and 59/100 for Argon high; the author also says Argon compacted context surprisingly fast and didn't feel like a 1M window.

A joking post says Meta being accused of gaming benchmarks may have given Google the confidence to release Argon 4 publicly.

A Chinese user says a single plain-text message used 15% of the five-hour Argon quota, and six messages locked them out for five hours.

A self-made cycling AI news radio episode leads with Google shipping Gemini 4 Argon, plus OpenAI and Anthropic on UK frontier-AI rules and AI protein drugs.

A user complains that Gemini 4 Argon in Antigravity is available only to Ultra subscribers, leaving lower tiers without access.

A day-long test finds 3D game scenes weak, image generation unremarkable, and 10-second chat-generated video the highlight, while the five-hour quota and slower mini-game builds lag ChatGPT.

A meme post jokes that Gemini 4 Argon appears and disappears for Pro users inside Google Antigravity, making it hard to use reliably.