GPT-6 Astra Drives a Real Corolla Through a Cone Course

OpenAI's GPT-6 Astra finished a 134.7-meter parking-lot cone course behind the wheel of a modified Toyota Corolla on Sept. 23, becoming the first model tested by DrivingBench to complete the run. The model needed a second attempt, covering just 49 percent of the course on its first try before finishing in 5 minutes 22 seconds on the second.

DrivingBench is a benchmark built by three engineers: Aditya Ramabadran, a mathematics and computer science student at UC Berkeley, along with Simon Mahns and Tobias Gessler. The team wired a 2022 Corolla with comma.ai's comma four hardware running the openpilot driver-assistance stack, exposing three tools, observe(), set_motion() and stop_now(), that let a model watch the road and issue driving commands one at a time inside a single chat session.

The course and its rules

The route runs roughly 130 meters through a field of traffic cones. A model scores by how far it travels along the centerline before straying more than 4 meters from it, with a speed cap of 3.5 meters per second and up to three attempts per model. GPT-6 Astra is the only one of four tested models to reach 100 percent completion. Anthropic's Claude Fable 5.1 reached 45 percent on its third attempt, xAI's Grok 4.6 reached 11 percent, and OpenAI's own GPT-5.6 Sol stalled at 6 percent on all three tries without improving.

1. GPT-6 Astra100 %
2. Claude Fable 5.145 %
3. Grok 4.611 %
4. GPT-5.6 Sol6 %

Source: DrivingBench leaderboard, accessed Sept. 24, 2026

What the winning run cost

DrivingBench's trace for Astra's second attempt logs 6.6 million tokens and $7.74 in API charges for the 5-minute, 22-second drive. The site does not publish a token or dollar figure for any of the three models that failed to finish, and it gives no comparison to the cost of the gasoline a human driver would burn over the same 134.7 meters.

Astra balked before it drove

Astra "sometimes refused to drive the physical car, citing safety" earlier in testing, according to OfficeChai's account of the project, until the DrivingBench team renamed the tool server "DrivingBench Sandbox." The change altered a label the model reads in its own tool list, not anything about the car, the brakes or the course.

What the test does not measure

DrivingBench says every model got a single evaluation, with all three attempts run inside one continuous conversation, so a model could adjust after failing. A human safety operator stayed ready on the brake throughout, the course cameras have a narrow field of view, and the steering calibration was tuned conservatively, according to Trending Topics' report on the test. Those constraints mean the benchmark measures tool-calling and spatial reasoning under one specific setup rather than driving competence of the kind judged by sustained deployments, like the warehouse runs behind Agility Robotics' Digit. DrivingBench's site does not say whether it plans to test more models or open the course to outside submissions.