Gemini 4 Argon Built the Eiffel Tower and Paris in 3D From 8 Images

Gemini 4 Argon built a walkable 3D Paris around the Eiffel Tower in one night: the tower modeled in Blender, the city in Three.js, people, pigeons, a night mode and a day to night time-lapse, all running in a web browser. The bill was $48.97 for 3 h 24 min of agent work, from 8 reference images and 159 messages (measured from the OpenCode session log, 15 October 2026). This post is the making-of, with my real prompts, typos included.

Explore it free: kartit.net/games/paris-3d. Watch the 2-minute making-of: on YouTube.

Key facts (quick answer)

Fact	Paris 3D by Gemini 4 Argon
Date	15 October 2026, 00:11 to 07:48
Billed for Gemini 4 Argon	$48.97
Agent working time	3 h 24 min
My messages	159 (33 with pictures)
Reference sheets at the start	8
Tools	OpenCode + Blender (tower), Three.js (city)
Music	Lyria 3 Pro, one track, $0.08
Runs on	Desktop browsers and iPhone

TL;DR

• 8 reference sheets in, a walkable 3D Paris out, in one night.
• $48.97 billed, 3 h 24 min of Argon working.
• 159 messages from me, 33 with pictures: this was directed, not one-shot.
• First try: the whole city inside a giant ball.
• The tower took the most rounds: thickness, curve direction, antenna length.
• The music was the hardest part; the track you hear cost $0.08.

How did it start?

With a giant ball. I first asked for a 3D walk near Châtelet, then switched to the Eiffel Tower. Seventeen minutes later Argon had a first Blender scene: the tower inside a sphere, a street photo wrapped around the whole city. My next message was three words: “remove this ball”.

How did Argon get the Eiffel Tower right?

With reference sheets and corrections. I gave it the tower from every side, the top up close, and a tree. Then I pointed at what was wrong and Argon fixed it, rendered, and checked again. Some of my real messages:

• “effiel thower top part look thick”
• “unlike image shiwh is mor inward curvaed / this more like outward curved”
• “the tour does not have all that long antena like check the refernces images and iterate till prfect match”

How did it build the city?

From Google Earth views. The early city had too many blocks around the tower (“too mubch building near effiel tour that is not true”) and no river (“where is la seine”). I sent views from Google Earth; Argon moved the Seine, opened up the Champ de Mars, and rebuilt the blocks as four kinds of Haussmann buildings.

What about people and the night?

People walking, bikes and pigeons came next (“fix people no more cube cirl heads”, “pigeons instead of white birds”). Then the night: golden lights, the sparkle and the beam on top, plus a day to night time-lapse.

What was the hardest part?

The music. One version had too much accordion, another sounded Greek. The track in the final page is one Lyria 3 Pro generation that cost $0.08.

Caveats

• This is one project on one night, not a benchmark. The cost depends on how many rounds of corrections you need.
• The 159 messages matter: the result came from directing the model with pictures and short corrections, not from a single prompt.
• Google Earth views were used as references only; they are not part of the page.

My take

What impressed me was not the first try (a ball) but how well Argon took visual corrections. A screenshot plus five words like “top part look thick” was enough for it to change the right part of the model. That makes a 3D scene something you can art-direct in plain language.

Sources

• OpenCode session log, sessions ses_ec387278 and ses_ec8d3e31, 15 October 2026 (cost, times, message counts).
• Playable page: https://www.kartit.net/games/paris-3d/
• Making-of video: https://youtu.be/y57aBW6ICEA

Independent project by Anass Kartit. Not affiliated with Google.