AI agent drives a Toyota to a burger joint: only GPT-6 Astra finishes the route in DrivingBench

In a Sep 17 test, GPT-6 Astra in Codex drove a Toyota Corolla to In-N-Out on its second attempt.

GPT-6 Astra completed the route in 5 minutes 22 seconds and sent the car 24 commands. The authors calculated the cost of the successful attempt using published rates: $7.74.

DrivingBench connects a car via comma/openpilot to apps running coding agents. Engineers Aditya Ramabadran, Simon Mans and Tobias Gessler connected Codex, Claude Code and Cursor via MCP and tested four models.

Driving results. Each model was given up to three attempts in a single chat with context preserved. The prompts for analyzing errors were identical, so subsequent attempts took previous ones into account.

| Model and app | Route completed by attempt | | --- | --- | | GPT-6 Astra in Codex, reasoning medium | Completed the route on the second attempt | | Claude Fable 5.1 in Claude Code | 9%, 10%, 45% | | Grok 4.6 in Cursor | 8%, 11%, 10% | | GPT-5.6 Sol in Codex | 6%, 6%, 6% |

The authors tested the system on a 2022 Toyota Corolla. Each model had one series of dependent attempts.

The commands had to change. Initially, the models specified path curvature and distance in meters, but estimated distances poorly. The authors switched to speed, movement duration and steering position via `steering_percent` from 0 to 100%. The models performed better with this interface.

The authors released DrivingBench Harness v1.0 for connecting the apps. Laptop installation targets macOS and requires Python 3.12 and uv. After cloning the repository, run these commands:

```sh uv sync --locked uv run drivingbench install --client codex --client claude --client cursor --host comma ```

Then restart the apps and open the local dashboard at `http://127.0.0.1:8766`. The car requires comma four and a compatible Toyota on which openpilot controls acceleration and braking. The test used openpilot v0.11.2, commit `76a7f857`.

The system is installed separately on comma: run `uv run drivingbench bundle`, transfer the build and launch `activate.py` on the device while the car is parked. Installation must be followed by steering calibration and a check with a driver.

Sources: [WIRED, Oct 7, 2026](https://www.wired.com/story/ai-is-driving-cars-now-oh-boy), the DrivingBench README and results table dated Sep 17, and the authors' arXiv paper dated Sep 30.

The authors published test recordings from two cameras, commands, model explanations, telemetry and transcripts on Hugging Face under CC BY 4.0.

AI agent drives a Toyota to a burger joint: only GPT-6 Astra finishes the route in DrivingBench

In a Sep 17 test, GPT-6 Astra in Codex drove a Toyota Corolla to In-N-Out on its second attempt.

GPT-6 Astra completed the route in 5 minutes 22 seconds and sent the car 24 commands. The authors calculated the cost of the successful attempt using published rates: $7.74.

The authors tested the system on a 2022 Toyota Corolla. Each model had one series of dependent attempts.

The authors released DrivingBench Harness v1.0 for connecting the apps. Laptop installation targets macOS and requires Python 3.12 and uv. After cloning the repository, run these commands:

```sh uv sync --locked uv run drivingbench install --client codex --client claude --client cursor --host comma ```

Sources: [WIRED, Oct 7, 2026](https://www.wired.com/story/ai-is-driving-cars-now-oh-boy), the DrivingBench README and results table dated Sep 17, and the authors' arXiv paper dated Sep 30.

The authors published test recordings from two cameras, commands, model explanations, telemetry and transcripts on Hugging Face under CC BY 4.0.