DrivingBench Evaluates Multimodal LLMs Like GPT-6 Astra in Autonomous Driving Tasks

The new DrivingBench benchmark evaluates frontier multimodal vision-LLMs like GPT-6 Astra on autonomous driving tasks, testing their ability to translate vision feeds into real-time control actions. This demonstrates that general-purpose vision models are gaining spatial reasoning capable of handling complex control tasks, potentially shifting self-driving technology from modular engineering to unified end-to-end neural networks. However, severe latency constraints currently prevent cloud-based LLMs from being safely deployed in real-world vehicles. DrivingBench tests models on driving simulations and real-world vehicles equipped with open-source control hardware like comma.ai. While models like Astra showcase strong spatial benchmark performance, high inference latency remains the primary blocker for real-time actuation.