GPT-6 Astra has gained the ability to drive a car

DrivingBench is an experimental evaluation platform that gives large language models direct control of a real Toyota Corolla’s steering, accelerator, and brakes and measures how well they navigate a fixed cone course. Agents run up to three attempts in a single continuous chat; accepted actuator commands (set_motion, stop_now) and trajectory traces are logged and replayable alongside anonymized video. Performance metrics include percent of course centerline traversed while staying within 4 m, integrated GPS distance, finish time for completed runs, collisions that terminate attempts, token usage, and estimated API cost. The interface shows per-attempt progress, accepted commands, and replays, and frames are blurred for privacy. The project is research software and not affiliated with vehicle or model vendors.

Results show large variation across frontier models. GPT-6 Astra achieved a complete run (100% progress) in 5:22 on its second attempt, after a first attempt that reached 49%, using substantial token counts and costing several dollars; other models like Claude Fable 5.1 reached up to 45% before failing, Grok 4.6 and GPT-5.6 Sol managed single-digit to low-double-digit progress and multiple DNFs. The leaderboard highlights that a top-tier multimodal model can successfully drive the instrumented car on the cone course under these constrained evaluation rules, while many contemporary models still fail to complete the track.