GPT-6 Astra is the only model to finish a real-car cone course
Three independent researchers let Astra control a Toyota Corolla's steering, accelerator and brake via MCP; it completed the course on its second attempt at 1.5 km/h, while GPT-5.6, Grok 4.6 and Claude Fable 5.1 all failed.
DrivingBench project where 3 independent researchers tested AI driving on OpenAI's GPT-6 Astra model, the test was performed on a Toyota Corolla at an open area there was no crowd, steering, accelerator and brake controls were given to the Astra though it drove the car at a very low pace of 1.5 km/h which is even slower than walking speed and it passed the test on its 2nd attempt, it was the only model to complete the full course while GPT-5.6, Grok 4.6 and Claude Fable 5.1 all failed Astra even refused to drive at times for safety reasons, and the run cost about $92 per mile in tokens, roughly 500x the price of fuel. Note the speed was capped very low by the researchers themselves still AI has a long way to go in driving first it has to improve driving skills in crowded areas then has to improve speed and a lot of other factors as well