Test: Gemini 4 Argon built a racing game from one desk photo

A developer gave Gemini 4 Argon a single photo of a stranger's desk and two hours with no help: the model built a recognizable track with a ramp onto the shelf, a kart, engine sound and a lap counter. Its own test script, lacking a timeout, got stuck for 22 minutes, and one nudge fixed it in 13 minutes. The run used about 6.66 million tokens and cost roughly $3.

• The model cut 10 textures from the photo and built the desk and kart via Blender scripts
• The track climbs a ramp onto the shelf and comes back down
• A test without a timeout stalled for 22 minutes; the fix took 13 minutes
• About 6.66 million tokens used, with the run costing roughly $3