TLDR;
Google opened Gemini for Argon to everyone on 13 October. The video reviews ten creations made with it, from a fluffy pumpkin and a playable synthesiser to a 3D chess game and a pelican riding a bicycle. The strongest results usually came from repeated cycles of building, reviewing and fixing; comparisons using a single prompt showed weaker physics and fine detail. The video also notes Argon’s cost, where the projects can be tried, and that launch-week demonstrations may represent the best-case results.
- Argon can produce varied interactive projects, including games, artwork and 3D scenes.
- Repeated feedback and revision often make a bigger difference than a single prompt.
- Argon appears fast and relatively inexpensive in some tests, but can lag behind on realistic physics and fine details.
- Several projects include their prompts or code, making it possible for others to inspect or reproduce them.
Intro [0:00]
Google opened Gemini for Argon to everyone on 13 October. Within two days, social media was full of projects such as chess pieces, temples and a pelican riding a bicycle. The video selects ten creations and explains what each one is, how it was made and what makes it notable. Before covering the top three, it also compares a few tests in which Argon performed poorly.
#10: A Fluffy Pumpkin [0:26]
The developer known as the bugged dev asked Argon to make a plush pumpkin. Argon created it entirely in Three.js, a JavaScript 3D graphics library, without Blender or a separate model file. Its fur consists of 28 layered shells and 60,000 individual strands. A mouse-controlled air jet lets users blow the fur around, and the result still looks soft when viewed up close.
#9: A Synth You Can Play [0:53]
Alex Petrascu asked Argon to make a synthesiser combining elements of a Roland Jupiter and a Moog, and said the work took less than 30 minutes. The result is a single HTML file with eight voices, a ladder filter, a 16-step drum machine and a piano roll that can play dropped-in MIDI files. MIDI is a standard format for storing musical notes and performance instructions. Every sound is generated in code, with no audio samples. The synthesiser is available at argonsynth.netlify.app, where listeners can judge how closely it resembles the original hardware.
#8: An Apple Museum [1:27]
Justine Moore asked Argon to create an interactive museum of Apple computers without supplying images or specifications. Argon found archival photographs, built a tool to clean them up, gathered specifications spanning 50 years and made a 3D Cover Flow-style display featuring 87 Macs. The page’s visual style changes as visitors move between decades. The research and preparation of the material are especially significant, although the specifications should still be checked before being quoted.
#7: A Helicopter Made of Code [2:04]
A creator called B built a futuristic attack helicopter in Three.js, with every part generated from code. The scene includes the helicopter’s rotors and missiles, as well as a landing pad, desert and snowy forest beneath the northern lights. B said the project took several rounds of work between Argon and Nano Banana, Google’s image-generation model. It was not made with one prompt: the process involved building the scene, reviewing it and making further changes, a pattern seen in many of the strongest projects.
#6: Six Games, Code Included [2:38]
Ammar Reshi, who leads product and design at Google AI Studio, created six games and shared their prompts and code on GitHub. They include Argon Tennis, where a phone acts as the racket; Argon Valley, a cosy island-life simulation; Grand Theft Argon, an open-world game with cell-shaded graphics and an 1980s look; and The Corridor, a VHS-style horror game that Reshi said startled his own team. He said each game came from a single shot. Since the prompts are public, others can rerun them and check how the projects were made.
#5: 534 Paintings, All Code [3:16]
Surya, a designer at WebSim, shared 534 paintings made entirely from code, along with the code for each one. The images include a rooster with individually drawn feathers, a bowling ball rolling towards the viewer and a bicycle in the snow at a railway crossing. Some resemble old prints, while others look like glitch art. Not all 534 are equally successful, but several are considered good enough to display as wall art.
#4: Desk Racer [3:45]
Genevieve, who works in developer relations for Google AI Studio, photographed a home-office desk and asked Argon to turn it into a racing game. The track passes a glowing keyboard, loops around a monitor arm and runs by a small cat figure. Argon modelled the desk in Blender through MCP, or Model Context Protocol, while Nano Banana created the textures and Three.js runs the game. Genevieve said it was her first serious 3D project. The game can be driven in AI Studio.
Where Argon Lost [4:19]
In a gummy-bear test that had more than 600,000 views, Semchenko gave Opus 5.5 and Argon the same prompt, then subjected both creations to the same actions: pulling a paw, flattening a bear, throwing it and pouring a bag of bears over it. The Opus bears stretched and wobbled like jelly, while Argon’s were stiff and were described as “gummy statues”. Argon also cost more in that test: $22 compared with about $7.
A separate firefly-jar test showed a similar difference in visual quality. Opus took 64 minutes and cost $15 to create realistic-looking glass with a tiny cabin inside. Argon took eight minutes and cost 19 cents, but the creator thought its result looked like a glow stick in a box. These comparisons suggest Argon can be fast and cheap but may struggle with physics and fine detail. West Boss, who tested it for a week before launch, also found it strong at coding and copying user interfaces, while noting that many results were impressive demonstrations rather than production-ready work.
#3: A Temple in Blender [5:21]
Sundar Pichai shared a temple complex built in Blender, including a five-storey pagoda, a red arched bridge, stone lanterns and a golden hall containing statues. One agent builds the scene while another examines the render and critiques it; they repeat the process until the result looks right. The next day, Ashutosh Srivastava used the demonstration video as input to Antigravity, worked out how it had been made and built another temple. His version was rougher, with many plain white shapes, but captured the layout. The independent recreation, completed in a day, makes the method particularly convincing.
#2: Wizard Chess [6:01]
Demis Hassabis posted a complete 3D chess game set in a stone hall lit by wall torches. The pieces carry swords and shields. Google said Argon wrote scripts to model each chess piece, used the same build-and-critique process to improve their appearance, added shatter physics for captures and created a computer opponent. Hassabis said the game reminded him of Battle Chess on the Amiga. It can be played in AI Studio.
#1: A Pelican on a Bicycle [6:30]
The most popular project, with more than two million views, shows a pelican riding a bicycle. Simon Willison has used the image as a recurring test for AI models. Faffer, who works at Google DeepMind, created it using Fourier epicycles: a chain of rotating circles, each following the previous one, with the final circle carrying a pen. Three circles produce a rough squiggle, but with 2,049 circles the error falls to one-third of a per cent and the drawing becomes a clear pelican. The pelican then pedals as the camera moves around it in 3D. Argon also made the video and wrote the music prompt. The project combines a silly subject with a mathematical technique, which helped it spread widely.
What the Results Show [7:15]
Five of the ten projects came from people working at Google, and companies tend to present their strongest demonstrations during launch week, so the examples may not represent typical results. Still, projects from outside Google, including the pumpkin, synthesiser and helicopter, also worked well. The most impressive outcomes often came from repeatedly asking Argon to build something, inspecting a render and correcting it, sometimes over several hours. Comparisons based on a single prompt showed its weaker side.
At the time of the video, Argon cost $2 per million input tokens and $10 per million output tokens, with those prices due to double after 30 November. Desk Racer and Wizard Chess were available in AI Studio, while Ammar Reshi’s games were on GitHub. Links to the projects were in the video description.