GPT-6 Astra has gained the ability to drive a car
3 hours ago
- DrivingBench evaluates frontier models on real-world driving by controlling a Toyota Corolla's steering, accelerator, and brakes on a fixed cone course.
- Models have up to 3 attempts in one continuous chat, with performance metrics including progress along the course centerline, distance, finish time, commands, and token cost.
- The leaderboard shows GPT-6 Astra (Codex) leading with 100% progress and a finish time of 5:22, followed by Claude Fable 5.1 (45%), Grok 4.6 (11%), and GPT-5.6 Sol (6%).