September 27, 2026
Code agents take the wheel: GPT-6 Astra completes a 134.7-meter course
On Sep 24, 2026, GPT-6 Astra completed the 134.7-meter course on its second attempt in 5:22. DrivingBench let the model control a real car through MCP.
Now DrivingBench connects Codex, Claude Code, and Cursor to a car through an MCP server with observation, movement, and emergency-stop commands.
On the same course, Claude Fable 5.1 reached 45%, Grok 4.6 reached 11%, and GPT-5.6 Sol reached 6%. None of the three models finished.
Not a garage setup. On a laptop, you need to clone the repository, run `uv sync --locked`, and install the clients with `uv run drivingbench install`, then restart the chat apps. You need macOS, Python 3.12, `uv`, SSH access to comma, and an MCP client with images and tool calls.
The operator opens a local dashboard, sets the cameras and speed limit, starts openpilot, and copies the ready-made prompt into the chat. The model chooses a speed from 0.5 to 3.5 m/s and a command duration from 5 to 60 seconds.
Astra's successful run used 6.6 million tokens and cost $7.74 at list price. Average speed was 0.94 mph.
The reference configuration used a comma four with openpilot v0.11.2 and a Toyota Corolla with torque steering on an empty private test site.
The MIT-licensed DrivingBench repository was published on Sep 24, 2026.
