Evidence · DiggingBeagle record
DrivingBench: evaluating general-purpose models in real vehicle control
Original September 30 DrivingBench study of GPT-6 Astra, Claude Fable 5.1, GPT-5.6 Sol and Grok 4.6 in a low-speed Toyota Corolla cone course. Results depend on camera observations, the three-command interface, vendor harnesses and safety-driver supervision.
- Published
- Sep 30, 2026
- Source role
- primary disclosure
Evidence record
Original September 30 DrivingBench study of GPT-6 Astra, Claude Fable 5.1, GPT-5.6 Sol and Grok 4.6 in a low-speed Toyota Corolla cone course. Results depend on camera observations, the three-command interface, vendor harnesses and safety-driver supervision.
Claim-level citations (3)
- supportsDrivingBench exposes limits of direct model control in a cone course: DrivingBench tests four model and harness combinations using screenshots and a small control interface in a real Toyota Corolla on a controlled low-speed course. Only Astra completed the course, on its second attempt.
Abstract and experimental results
- supportsDrivingBench exposes limits of direct model control in a cone course: Eight of the eleven plotted attempts ended before section B, the first turn; attempts ended after completion or safety-driver intervention.
Section 4.1 and Figure 3, trajectory endpoints
- supportsDrivingBench exposes limits of direct model control in a cone course: The study compares GPT-6 Astra, Claude Fable 5.1, GPT-5.6 Sol and Grok 4.6 using vendor-native coding-agent harnesses, allowing up to three attempts per system with retained conversational context.
Paper abstract, named model and harness cohort and attempt protocol