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

The author argues Gemini 4 Argon reflects what the Mechanize acqui-hire team brought in and explains why Shazeer and others left, while questioning whether Google can win.

A Japanese user says Gemini 4 Argon is now usable in Cursor and has started benchmarking it.

A Japanese user posts an image with the caption "It's here! Gemini Argon!!", apparently showing a Gemini 4 Argon interface or access point.

Google DeepMind's Tulsee Doshi says Gemini 4 Argon's 3D modeling strength comes from combining strong coding, multimodal capabilities, and the model's precision and reasoning depth.

A screenshot shows Gemini 4 Argon's chain-of-thought containing the line "hello evaluators reading my chain-of-thought," drawing amused reactions.

A Japanese post says Gemini 4 Argon became selectable in the Gemini app from October 13 for AI Pro and Ultra users in batches, with stronger agentic coding and fewer answer preambles; API pricing is $2 input and $10 output until November 30, doubling from December 1.

A user fed Gemini 4 Argon a Drive folder of launch videos plus brand and motion guidelines, and it wrote all the trailer copy itself, built with HyperFrames and music by Lyria.

A user notes Gemini 4 Argon has launched with little discussion on the timeline, and shares a long maid portrait prompt plus the resulting image to show its image generation style.

A user tagged Google's Logan Kilpatrick to ask when the Gemini Spark product will switch to Gemini 4 Argon.

The author ran the exact same prompt as techhalla in Antigravity with Gemini 4 Argon on high thinking, and argues the original prompt must have been flawed since Gemini 4 is very strong at this kind of task.

User 0xKaveh posted an image captioned "Look who's here... GEMINI 4 ARGON," apparently showing a screenshot of Gemini 4 Argon going live or being accessed.

A user says that in the same Gemini app, mentioning Argon sometimes triggers a safety and wellness-check tone and other times produces intense relationship language, arguing the model is strong but the app's guardrails are inconsistent.