A $13, 4B model trained with RL outflies GPT-6 Astra at landing a flying Tesla Roadster
Post-training 4B model with RL to fly Tesla Roadster 21% better than GPT-6 Astra
SkyPilot built a physics-simulator RL environment around Tesla's flying Roadster demo, with four air thrusters, gusty wind, a ramp flip and a kickflip over a Cybertruck. Every frontier model tested — DeepSeek V4.1 Flash, Muse Spark Open, Beam 1.0, Nemotron 3.5 Super, GPT-6 Astra — lands 0% of kickflips in real time, since they need 1 to 6 seconds per decision while the car demands one every 0.2 seconds. A Qwen3-4B model RL-trained for about $13 answers in 0.09 seconds and lands 90% in a storm.
When we paused the world so the models could think as long as they wanted, they still lost to a 4B model we trained with RL in a storm.